Information Revelation and Pandering in Elections∗

Navin Kartik† Francesco Squintani‡ Katrin Tinn§

April 7, 2015

(First version: March 2012)

Abstract

Does electoral competition promote efficient aggregation of politicians’ policy-relevant private

information? This paper suggests otherwise when politicians are (primarily) office-motivated. In

a two-candidate Hotelling-Downs framework, we establish that the electorate’s welfare cannot

be higher than what can be obtained by disregarding the information of one politician. Under

canonical information structures, politicians have an incentive to overreact to their information

relative to the electorate’s prior belief, i.e. to “anti-pander” rather than pander. We argue that

the information-aggregation inefficiency can be traced to the politicians’ office motivation; in a

suitable sense, it also applies to alternative game forms when politicians make non-binding policy

announcements.

∗For helpful comments and discussions we thank Nageeb Ali, Scott Ashworth, Kfir Eliaz, Andrea Galeotti, MatiasIaryczower, Alessandro Lizzeri, Massimo Morelli, David Rahman, Jesse Shapiro, Olivier Tercieux, several seminarand conference audiences, and, especially, Alfonso Goncalves da Silva, Johannes Horner, and Balazs Szentes. Kartikgratefully acknowledges financial support from the Sloan Foundation, the National Science Foundation (Grant SES-115593), and the hospitality of and funding from the University of Chicago Booth School of Business during a portionof this research.

†Columbia University, Department of Economics. Email: nkartik@columbia.edu.

‡University of Warwick, Department of Economics. Email: F.Squintani@warwick.ac.uk.

§Imperial College London, Business School. Email: k.tinn@imperial.ac.uk.

1. Introduction

A fundamental question for representative democracy is whether citizens are sufficiently informed to

select representatives whose policy choices will enhance their welfare. While citizens may themselves

be poorly informed on policy issues—as posited by Downs (1957) in his “rational ignorance” hypoth-

esis and since documented by numerous studies starting with Campbell, Converse, Miller and Stokes

(1960)—political candidates devote substantial resources and have broad access to policy experts

and think tanks. Information garnered by politicians may be conveyed to the electorate through

their policy positions and electoral campaigns.1 This paper examines whether elections provide the

appropriate incentives for politicians to efficiently reveal their policy-relevant (private) information.

If politicians were benevolent, it is plausible that there would be no strategic impediments. But

can efficiency be achieved when politicians are not benevolent? In particular, how well do elections

aggregate the information of office-seeking politicians?

One prevalent view is that elections function well because even office-seeking politicians are im-

pelled to choose policies that promote voters’ interests. Indeed, in an influential article titled “Why

Democracies Produce Efficient Results”, Wittman (1989, p. 1400) argued that political competition

benefits the electorate because “there are returns to an informed political entrepreneur from pro-

viding the information to the voters, winning office, and gaining the [...] rewards of holding office.”

Concurrently, however, there are also charges—both in popular circles and in some recent academic

work that we discuss subsequently—that competitive pressures drive office-motivated politicians to

pander to voters’ opinions. After all, the argument goes, it is hard to win an election by campaigning

on policies that voters view as unlikely to be good; a politician is better off just promising to do

whatever voters are inclined towards. Pandering is viewed as inefficient because it would lead to

policies that are excessively distorted toward the voters’ less-informed opinions.

In this paper, we (re-)assess the efficiency of elections when office-seeking politicians possess

policy-relevant private information. We begin in Section 2 by proposing a natural extension of

the classic model of representative democracy: the two-candidate Downsian model (Downs (1957),

building on Hotelling (1929)). We maintain all the usual assumptions, including that politicians

1 There is evidence that voters learn, update, or refine their views during elections. See, for example, experimentson deliberative polling by Fishkin (1997); empirical studies on the effects of information on voters’ opinions by Zaller(1992), Althaus (1998) and Gilens (2001); studies on framing in polls such as the ones by Schuman and Presser (1981);and experiments on priming by Iyengar and Kinder (1987).

1

maximize the probability of winning the election and make policy commitments, but introduce one

twist: each of the politicians has imperfect private information about policy consequences. In other

words, each politician has socially-valuable information about which policy would be best for a

representative or median voter. For concreteness, we assume a familiar normal-normal information

structure;2 as we discuss, our results can be generalized to other statistical structures.

We establish in Section 3 that Downsian elections cannot efficiently aggregate both politicians’

private information. Theorem 1 identifies a minimum degree of inefficiency: the voter’s welfare

in any electoral equilibrium is equivalent to policy being implemented based on just one politician’s

information, while not necessarily using even this information efficiently. Consequently, voter welfare

is no higher than what can be obtained by disregarding the presence of one politician entirely. This

inefficiency is general—across statistical structures and voter preferences—and we explain how it can

be traced to two implications of the politicians’ office motivation: (i) as far as they are concerned,

the politicians are engaged in a constant-sum game; and (ii) even though each of them has socially

valuable information, their information does not directly affect the politicians’ payoffs.

Why isn’t it an equilibrium for each politician to propose a policy that is best for the voter

based on his own information, i.e. to use an “unbiased strategy” (in which case the election would

aggregate more than one politician’s information)? Proposition 1 shows that, perhaps contrary to

intuition, politicians would have an incentive to deviate by “anti-pandering”—overreacting to their

private information—as the rational voter would elect the more extreme politician under unbiased

strategies. The voter would do so because each politician’s estimate of the state based on his own

signal puts more weight on the prior than the voter’s estimate after learning both politicians’ signals.3

Building on the above logic, we identify in Proposition 2 a symmetric equilibrium that features

anti-pandering by both politicians. In this equilibrium, politicians choose different platforms with

probability one. The anti-pandering mechanism provides a new perspective on the classic issue of pol-

icy divergence. Unlike some other prevalent explanations (e.g., ideologically-motivated candidates as

2 More precisely: the best policy for the voter—hereafter, the “state” (of the world)—is drawn from a normaldistribution; each candidate’s private signal is the true state plus noise that is also normally distributed. The voter’spayoff is given by a quadratic loss function of the distance between the chosen policy and the state.

3 Glaeser and Sunstein (2009) and Roux and Sobel (2013) also identify this implication of Bayesian updating in a non-strategic group decision-making context. The underlying statistical property holds whenever the posterior distributionof the state given a politician’s signal is in an exponential family and the prior is conjugate with the posterior; thisclass includes a variety of familiar information structures as elaborated in Subsection 4.1.

2

in Wittman (1983) and Calvert (1985)), anti-pandering features office-motivated politicians diverg-

ing in order to increase their support from a median voter whose ideology is known.4 There is some

empirical evidence that politicians can secure larger vote shares by adopting extreme platforms; see,

for example, Stone and Simas (2010) on U.S. House elections. The anti-pandering mechanism even

allows for each politician’s platform to be more extreme than the preferred policy of the average voter

who votes for him, as we formalize by explicitly considering an electorate with heterogenous ideolo-

gies (Corollary 1). In this quite stringent sense, political competition can lead to “over-divergence”

or policy polarization despite politicians solely seeking office. Although it would be premature to

directly link the mechanism we explore here with empirically observed policy divergence, it is inter-

esting to note that over-divergence in the aforementioned sense has been documented (Bruhn and

Greene, 2007; Sørensen, 2012). More fundamentally, anti-pandering qualifies the general notion that

office motivation produces a centripetal force toward the electorate’s median voter.

As with much of the literature, the bulk of our analysis assumes that candidates make commit-

ments to the policies they would implement if elected.5 From a positive point of view, it seems

reasonable to suppose that some degree of electoral commitment is available; in their meta-study

of earlier research, Petry and Collette (2009) conclude that around 67% of campaign promises have

historically been kept. The theoretical literature has proposed multiple rationales for commitment,

most prominently that of re-election concerns (Alesina, 1988).6 Nevertheless, particularly for our

normative conclusions, it is important to assess the role of the commitment assumption. Section 4

addresses this and other robustness issues. In a nutshell, we propose that incorporating non-binding

or “cheap talk” announcements—be they in lieu of commitment, preceding commitment, or as an

additional alternative to commitment—does not fundamentally alter our welfare conclusions. An

important caveat, however, is that this is now subject to selecting among multiple equilibria in a

manner that we explain and argue in favor of. The intuition for why cheap talk should be ineffective

4 Other explanations for divergence include those based on increasing turnout (Glaeser, Ponzetto and Shapiro, 2005),campaign contributions (Alesina and Holden, 2008; Campante, 2011), valence asymmetries (Groseclose, 2001; Aragonesand Palfrey, 2002), signaling character, competence, or related mechanisms (Callander and Wilkie, 2007; Kartik andMcAfee, 2007; Honryo, 2014), or the presence of more than two parties (Palfrey, 1984).

5 Exceptions include Besley and Coate (1997), Osborne and Slivinski (1996), and subsequent work using citizen-candidate models.

6 Another rationale, particularly relevant in our context, is that if there is uncertainty about a candidate’s qualityof information and candidates have reputation concerns (perhaps because of re-election motives), then “flip flopping”or “vacillating” may be associated with poor quality information, resulting in stickiness akin to commitment; see, forexample, Prendergast and Stole (1996) or Majumdar and Mukand (2004).

3

in our setting is as follows: if one politician discloses policy-relevant information while the other does

not, the latter is better informed and hence more attractive to the voter if policies have not already

been committed to. Thus, politicians prefer not to reveal any information through cheap talk, which

renders equilibria in which they do implausible (cf. Farrell, 1993).

In summary: our analysis suggests that one should neither expect (i) electoral competition be-

tween office-motivated politicians to result in policies that efficiently aggregate their private infor-

mation, nor (ii) policies to be distorted toward the electorate’s prior. Rather, while politicians’

private information can be revealed through their electoral platforms, it can result in anti-pandering.

More broadly, office motivation has stringent welfare consequences in our setting. Arguably, when

politicians possess private information of significant social value, a well-functioning representative

democracy demands either principled or benevolent politicians or alternative institutional arrange-

ments to majoritarian elections.

Related Literature. This paper ties most closely into the literature on electoral competition when

candidates have policy-relevant private information.7 Heidhues and Lagerlof (2003) illustrate why

candidates may have an incentive to pander to the electorate’s prior belief; their setting is one with

binary policies, binary states, and binary signals. We find that in our richer setting, precisely the

opposite is true for a broad class of information structures. Plainly, with binary policies, one cannot

see the logic of why and how candidates may wish to overreact to private information. Loertscher

(2012) maintains the binary signal and state structure, but introduces a continuum policy space.

His results are more nuanced, but at least when signals are sufficiently precise, the conclusions are

similar to Heidhues and Lagerlof (2003).8

Laslier and Van de Straeten (2004) show that if voters in the Heidhues and Lagerlof (2003) model

are endowed with sufficiently precise private information about the policy-relevant state, then there

are equilibria in which candidates fully reveal their private information; see also Klumpp (2014) and

Gratton (2014). By contrast, we are interested in settings in which any information voters possess

that candidates do not is relatively imprecise. While we make the extreme assumption that voters

7 There are contexts where candidates have private information that is not policy relevant for voters. For instance,the strategic effects of private information about the location of the median voter are studied by Chan (2001), Ottavianiand Sorensen (2006), and Bernhardt, Duggan and Squintani (2007, 2009).

8 Appendix E shows how overreaction or anti-pandering arises in a binary-signal example when the policies and thestate lie in the unit interval. This example is a special case of the statistical family mentioned in fn. 3 and discussedin Subsection 4.1; it permits a closer comparison with Heidhues and Lagerlof (2003) and Loertscher (2012).

4

have no private information, our main points are robust to (small) variations on this dimension.

Schultz (1996) studies a model in which two candidates are perfectly informed about the policy-

relevant state but are ideologically motivated. He finds that when the candidates’ ideological prefer-

ences are sufficiently extreme, platforms cannot reveal the true state; however, because of the perfect

information assumption, full revelation can be sustained when ideological preferences are not too

extreme. Martinelli (2001) and Martinelli and Matsui (2002) derive further results with ideologically

motivated candidates who are perfectly informed about a policy-relevant variable.

There are other models of politics in which policy distortions arise because office-motivated

politicians wish to influence voters’ beliefs; the mechanisms, however, are less related to ours because

these models are generally of a single incumbent with “career concerns” to build reputation for

either competence or aligned preference.9 While most of these papers focus on pandering toward the

electorate’s prior, anti-pandering arises in Prendergast and Stole (1996) and Levy (2004).

The literature on cheap talk in elections is limited. Harrington (1992, 1993b), Panova (2009),

Schnakenberg (2014), and Kartik and van Weelden (2015) identify different reasons why informative

communication is possible when politicians’ ideological preferences are their private information.

2. A Downsian Model with Expert Politicians

In our baseline model, an electorate is represented in reduced-form by a single (median or repre-

sentative) voter, whose preferences depend upon policy, y ∈ R, and an unknown state of the world,

θ ∈ R. Subsequently, we will explicitly discuss heterogeneous voter ideologies. We assume the voter’s

preferences can be represented by a von Neumann-Morgenstern utility function, U(y, θ) = −(y−θ)2.

The state θ is drawn from a normal distribution with mean 0 and finite precision α > 0 (i.e. variance

1/α). Each of two candidates, A and B, maximizes the probability of winning the election. Each

candidate i privately observes a signal θi = θ+ εi, where each εi is drawn independently of any other

random variable from a normal distribution with mean 0 and finite precision β > 0.10

9 Harrington (1993a) and Cukierman and Tommasi (1998) are early contributions; see also Canes-Wrone, Herronand Shotts (2001), Maskin and Tirole (2004), Prat (2005), and more recently, Morelli and van Weelden (2013) andAcemoglu, Egorov and Sonin (2013).

10 The normal-normal structure is a well-known family of conjugate distributions; Subsection 4.1 discusses otherstatistical structures. The assumption that both candidates receive equally precise signals is for expositional simplicity;our results also hold when one candidate is known to receive a more precise signal than the other. Such asymmetric“competence” can serve to capture incumbency advantage. Our results can also be extended in other directions,

5

After privately observing their signals, candidates simultaneously choose platforms, yA ∈ R and

yB ∈ R respectively. Upon observing both platforms, the voter updates her belief about the state and

then elects one of the two candidates. The elected candidate implements his platform as policy, i.e.,

platforms are commitments in the Downsian tradition. All aspects of the model except candidates’

privately observed signals are common knowledge, and players are expected-utility maximizers.

Equilibrium and welfare. With some abuse of notation, a pure strategy for a candidate i will

be denoted as a (measurable) function yi(·) : R→ R, so that yi(θi) is the platform chosen by i when

his signal is θi.11 A mixed strategy for the voter is a (measurable) function p(·) : R2 → [0, 1], where

p(yA, yB) represents the probability with which candidate A is elected when the platforms are yA and

yB. We are interested in (weak) perfect Bayesian equilibria of the electoral game (including those in

which candidates play mixed strategies), which implies that the voter elects candidate i if yi is strictly

preferred to y−i, where the subscript −i refers to candidate i’s opponent. As is common, we require

that the voter randomize with equal probability between the two candidates if she is indifferent

between yA and yB. This tie-breaking specification is not essential (see, in particular, fn. 17); but

it simplifies matters by pinning down voter behavior on any equilibrium path. Furthermore, for

technical reasons, we restrict attention to equilibria in which for any given policy of one candidate,

say yA, the voting function p(yA, ·) has at most a countable number of discontinuities, and analogously

for p(·, yB) for any yB.

The notion of welfare we use is the voter’s ex-ante expected utility.

Terminology and preliminaries. A pure strategy yi(·) is informative if it is not constant and it

is fully revealing if it is a one-to-one function, i.e. if the candidate’s signal can be inferred from his

platform. As is well known (Degroot, 1970), the normal-normal information structure implies that

the expected value of the state θ given a single signal θi is

E [θ|θi] =β

α+ βθi, (1)

including candidates’ signals being correlated conditional on the state and adding uncertainty about the precision oftheir signals.

11 A mixed strategy for candidate i is a mapping σi : R→ ∆(R), where ∆(R) denotes the set of probability measureson R. All our formal results and proofs cover mixed strategies for candidates; however, since candidates’ mixing doesnot play an essential role, the main text focuses on candidates using pure strategies for simplicity.

6

whereas conditional on both signals, the expected value is

E [θ|θA, θB] =2β

α+ 2β

(θA + θB

2

). (2)

Because of quadratic-loss utility, the optimal policy for the voter is the conditional expectation of

the state given all available information. Since the only information a candidate has when he selects

his platform is his own signal, we refer to the strategy yi (θi) = E [θ|θi] = βα+β θi as the unbiased

strategy. Plainly, this strategy is fully revealing. We say that a strategy yi(·) displays pandering to

the voter’s beliefs if for all θi 6= 0, yi(θi) is in between 0 and E[θ|θi], i.e. signθi = signyi(θi) =

signE[θ|θi]− yi(θi). In other words, a candidate panders if his platform is systematically distorted

from his unbiased estimate of the best policy toward the voter’s prior expectation of the best policy.

Analogously, we say that yi(·) displays anti-pandering or overreaction to private information if for

all θi 6= 0, signθi = signyi(θi) = signyi(θi)−E[θ|θi]. We say that a platform y is more extreme

than platform y′ if the former is further from the prior mean, 0, than the latter, i.e. if |y| > |y′|.

An equilibrium is informative if at least one candidate plays an informative strategy. Observe

that for any uninformative pure strategy, there is an equilibrium in which both candidates use that

pure strategy due to the latitude in specifying off-path beliefs. Our interest will be in informative

equilibria. An equilibrium is fully revealing if both candidates’ strategies are fully revealing. An

equilibrium is symmetric if both candidates use the same strategy. An equilibrium is competitive if

both candidates have an ex-ante positive probability of winning; it is non-competitive if one candidate

wins with ex-ante probability one.12

3. Main Results

3.1. Anti-Pandering and Polarization

Given that the voter desires policies as close as possible to the true state, and that a candidate’s only

information when choosing his policy is his private signal, one might conjecture that a candidate can

do no better than playing an unbiased strategy, particularly if the opponent is also using an unbiased

strategy. However:

12 A technical note is that because of the continuum of candidates’ signals and policies, various statements in theanalysis and proofs should be understood to hold subject to “almost all” qualifiers (indeed, even the definition ofequilibrium should be understood as imposing optimality of each candidate’s behavior for almost all signals); wesupress such caveats unless essential.

7

Proposition 1. The profile of unbiased strategies is not an equilibrium: candidates would deviate

by overreacting to their information.

(Proofs of this and all other formal results are in Appendix B.)

The incentive to overreact arises because if both candidates were to play unbiased strategies, the

voter would optimally select the candidate with a more extreme platform, i.e. the candidate whose

platform is larger in absolute value. Why? Since unbiased strategies are fully revealing, the voter

would infer both candidates’ signals from their platforms. Equations (1) and (2) imply that she

would form a posterior expectation that has the same sign as the average of the two candidates’

individual posterior expectations but that is more extreme.13 Since the candidate whose platform

is closer to the voter’s posterior expectation is elected, it follows that the voter would elect i if and

only if |yi| > |y−i|. Hence, each candidate would like to raise his probability of playing the more

extreme platform, which can be achieved by placing more weight on his private signal than what is

prescribed by the unbiased strategy. Consequently, overreacting to private information constitutes a

profitable deviation, whereas pandering does not.14

Despite the incentive to overreact, can information be revealed in equilibrium? Perhaps surpris-

ingly, we find that an appropriate degree of overreaction can support full revelation of information.

Proposition 2. There is a symmetric and fully-revealing equilibrium with overreaction where both

candidates play

y (θi) = E[θ|θi, θ−i = θi] =2β

α+ 2βθi. (3)

In this equilibrium, each candidate is elected with probability 1/2 regardless of the signal realizations

θA and θB. Furthermore, this is the unique symmetric pure-strategy equilibrium in which candidates

use fully-revealing and continuous strategies.

Not only does Proposition 2 establish existence of a fully revealing equilibrium, but it also provides

13 Indeed, the voter’s posterior mean can be larger in magnitude than both candidates’ platforms (rather than justtheir average); this is always the case when θA and θB are sufficiently close, since the posterior mean is continuous insignals, and for any θ, |E[θ|θA = θB = θ]| = 2β

α+2β|θ| > |E[θ|θA = θ]| = β

α+β|θ|.

14 Another way to understand the point is to note that when the voter conjectures unbiased strategies by thecandidates, equation (2) implies that for any platform of candidate −i, an increase in candidate i’s platform by ε > 0

would raise the voter’s posterior by 2βα+2β

(α+ββ

ε2

)> ε

2. Thus, increasing (resp., decreasing) his platform would benefit

candidate i when he is located to the right (resp., the left) of −i. Since under unbiased strategies a candidate’sexpectation of his opponent’s platform is in between zero and his own unbiased platform, pandering would reduce theprobability of winning while overreacting would increase it.

8

a sense in which uniqueness obtains. Indeed, Proposition 4 in Appendix B establishes a stronger

uniqueness result based only on “local” analogs of the properties required in Proposition 2.

The strategy given by (3) requires each candidate i to choose his platform to be the Bayesian

estimate of the state assuming his opponent has received the same signal. This is an overreaction

because he anticipates that, in expectation, his opponent’s signal will be more moderate than his

own, as the expectation of the opponent’s signal equals his unbiased estimate of the state, βα+β θi.

When the voter believes that both candidates overreact according to (3), platforms do not affect

winning probabilities because whenever candidate i increases his platform by δ > 0, the Bayesian

updating formula (2) implies that the voter’s posterior increases by 2βα+2β

(α+2β

2βδ2

)= δ/2, and thus

she remains indifferent between the two platforms.

Although the coefficient in (3), 2βα+2β , is increasing in the precision of candidates’ signals, β, and

decreasing in the precision of the prior, α, the same is true for the unbiased strategy’s coefficient,

βα+β . The degree of overreaction in the equilibrium of Proposition 2, as measured by 2β

α+2β −β

α+β , is

non-monotonic in the parameters: it is increasing in β (resp., decreasing in α) when β√

2 < α and

decreasing in β (resp., increasing in α) when β√

2 > α. The degree of overreaction vanishes as either

α or β tend to 0 or ∞. Thus, Proposition 2 predicts non-trivial overreaction when candidates are

somewhat better informed than the electorate but not overly so.

To further understand the degree of overreaction in Proposition 2, it is useful to briefly consider

heterogeneous voters. Suppose there is a continuum of voters; each voter k has utility function

Uk(y, θ) = − (y − θ − bk)2, where bk ∈ R parameterizes the voter’s ideology. The aggregate distri-

bution of bk is independent of all other random variables and given by a cumulative distribution

function F (·) that is continuous and symmetric around zero: for any b, F (b) = 1−F (−b). Thus, the

median ideology is zero and any voter k’s ex-ante preferred policy is her ideology bk. We say that

a voter k’s interim preferred policy is her ideal policy when she casts her ballot, E[θ|yA, yB] + bk.

To focus on the central issues, we assume that voters vote sincerely, i.e. for the candidate whose

platform is closest to their interim preferred policy. Plainly, the ordering of voters by their ideology

agrees with their ordering by either their ex-ante preferred policies or their interim preferred policies

no matter what strategies the candidates use. Whenever candidates adopt distinct platforms, say

yi > y−i, we say their platforms are polarized if yi is to the right of the interim preferred policy of

the average voter who votes for i, and y−i is to the left of the interim preferred policy of the average

9

voter who votes for −i.

Corollary 1. With a heterogenous electorate as just described, it remains an equilibrium for both

candidates to play (3); regardless of their platforms, both obtain a vote share of one half. Platforms

are polarized if and only if

2β

α+ 2β

|θi − θ−i|2

=|yi − y−i|

2> E[bk|bk > 0]. (4)

Since the anti-pandering strategies of Proposition 2 make the median voter indifferent between

the two candidates no matter their platforms, the equilibrium is preserved with a heterogeneous

electorate: now, all right-leaning voters (those with positive ideologies) will vote for the candidate

whose policy is further to the right, and conversely for all left-leaning voters. It bears emphasis

that anti-pandering is not driven by a desire to cater to the ideological distribution of voters; to the

contrary, the candidates’ equilibrium strategies are invariant to various properties of the ideological

distribution (subject to symmetry around 0).

Condition (4) identifies when candidates’ policies are polarized in the precise sense defined above.

The inequality says that the policies must be further apart than the ideology of the average right-

leaning voter.15 Such policies occur with positive ex-ante probability; whether the inequality is

satisfied or not depends on how disparate the candidates’ realized signals are. Ex ante, condition (4)

is more likely to hold when candidates are better informed relative to the electorate (higher β or lower

α) or when there is less ideological dispersion in the electorate.16 Overall, Corollary 1 establishes

that in the current model, the policies of the political elite can be quite polarized even when the

electorate’s ideological distribution is “nicely shaped”, e.g. single-peaked or otherwise concentrated

around the median.

3.2. Equilibrium Welfare

Return now to the baseline setting with a single (median) voter. Even though the equilibrium of

Proposition 2 fully reveals all available information to the voter, it does not use either politician’s in-

15 More generally, if F is not symmetric around its median, 0, then condition (4) can be written as12|yi − y−i| > maxE[bk|bk > 0],−E[bk|bk < 0].16 For example, if ideologies are normally distributed with mean zero, then E[bk|bk > 0] is increasing in the variance.

More generally, note that E[bk|bk > 0] = 2E[maxbk, 0] because F (0) = 1/2. As maxbk, 0 is a convex function, ithas higher expectation when F undergoes a second-order stochastically dominated shift, in particular when there is amean-preserving spread of F .

10

formation efficiently because of their overreaction. This prompts the question of what other equilibria

exist and whether there are any with higher voter welfare.

Fix an arbitrary equilibrium. Given the candidates’ (possibly mixed) strategies, denote by vi the

voter’s welfare (i.e. ex-ante expected utility) from electing candidate i no matter which policy pair

is actually proposed. Plainly, since this is a feasible strategy for the voter, the voter’s equilibrium

welfare cannot be smaller than maxvA, vB. Our key result, Theorem 1 below, proves that this is in

fact exactly the voter’s equilibrium welfare—even though, in equilibrium, the voter may be selecting

both candidates with ex-ante positive probability (as in the equilibrium of Proposition 2). Since a

candidate’s policy can only depend on his own information, it follows that it is as if the voter’s welfare

in any equilibrium is influenced by just one candidate’s signal and not by the other’s. Therefore,

electoral competition between office-motivated candidates leads to an inescapable inefficiency.

Theorem 1. In any equilibrium, there is some candidate i ∈ A,B such that the voter’s equilibrium

welfare is equal to the welfare obtained by electing candidate i no matter the proposed policies.

As the proof is somewhat involved but the result is central, we will sketch the main steps of the

argument. The key insight is Lemma 4 in Appendix B: any equilibrium must have the property that

for any on-path platform of a candidate i, say yi, the probability with which i would win when playing

yi cannot depend on what signal i has received. To establish this point, we first note that given any

strategy for the voter, p(yA, yB), the two candidates are engaged in a constant-sum Bayesian game

in which their payoffs depend only on the pair policy platforms and not (directly) on the signals. We

provide in Appendix A a general result about any two-player constant-sum Bayesian-game whose

payoffs depend only on the players’ actions and not on their types: every equilibrium has the property

that the distribution of actions played by some type of a player would also be a best response for

any other type of that player, even though the two types will generally hold different beliefs about

the opponent’s distribution of actions when types are correlated. A version of this result for finite

games has also been obtained by Viossat (2006).

Building on Lemma 4, we show in Lemma 5 that if an equilibrium is informative, it must be

an ex-post equilibrium in the sense that the voter’s strategy, p(yA, yB) must be constant across all

on-path platform pairs. Hence, a candidate would have no incentive to deviate even if he observed his

opponent’s platform before making his choice. Note, in particular, that this property is satisfied by

the equilibrium of Proposition 2. An intuition for this step is as follows. Consider any informative

11

equilibrium. Since at least one candidate’s platform is correlated with his signal, and since each

candidate’s signal provides him with information about his opponent’s signal, it follows that at least

one candidate’s signal is informative about his opponent’s platform. But then, the interim probability

of winning for a candidate can be independent of his signal (as required by Lemma 4) only if the

winning probability is independent of which platforms are played on the equilibrium path.17

This ex-post property for informative equilibria implies that in any equilibrium, informative or

uninformative, one of two cases holds: (a) the equilibrium is non-competitive and there is a candidate

i who is always elected regardless of the signal profile; or (b) the equilibrium is competitive and the

voter is indifferent between both candidates for all pairs of on-path platforms. In the latter case, the

voter’s expected utility given any on-path platform pair is, of course, the same regardless of whom

she elects. It follows that in either case, the voter’s ex-ante expected utility can be evaluated by

assuming that she always elects one of the two candidates, which is what Theorem 1 states.

Theorem 1 determines an upper bound on the voter’s welfare in any equilibrium of the Downsian

election because despite there being two informed candidates, the voter’s welfare might as well be

determined by only one of their signals and the corresponding policy platform. It follows that there

cannot be any equilibrium which delivers a higher welfare to the voter than she would obtain by

always electing one candidate who plays the unbiased strategy, i.e. who choses policy optimally based

on his signal. Our next result confirms that this outcome can be supported as a non-competitive

equilibrium, and furthermore, that any competitive equilibrium yields strictly lower welfare.

Theorem 2. There is a non-competitive equilibrium where the candidate who wins with probability

one plays the unbiased strategy. There is no equilibrium that yields the voter a higher ex-ante expected

utility, and any competitive equilibrium yields the voter strictly lower ex-ante expected utility.

There are multiple constructions to verify the first statement of Theorem 2. The simplest one is

such that the winning candidate, i, plays the unbiased strategy, the opponent plays an uninformative

strategy, say y−i(θ−i) = 0, and the voter believes that any deviation by candidate −i is also uninfor-

17 It is worth clarifying that this argument for informative equilibria does not use the assumption that the voter mustrandomize with equal probability when indifferent between candidates’ platforms. The ex-post property also applies touninformative equilibria, but only under the randomization assumption. If the voter were not required to randomizeuniformly when indifferent, there are uninformative equilibria with the flavor of “matching pennies”: for example, bothcandidates randomize uniformly over −x, x for some x > 0, and the voter elects candidate A if yA = yB while sheelects candidate B if yA = −yB (and randomizes with equal probability off the equilibrium path). Nevertheless, theconclusion of Theorem 1 applies to uninformative equilibria as well even without the randomization assumption.

12

mative. Interestingly, the same outcome can be supported also with −i playing the fully-revealing

strategy y−i(θ−i) = θ−i. Candidate −i is overreacting to his information here to such an extent,

choosing what would be an unbiased estimate only under a Laplacian or improper prior, that the

voter never finds it optimal to elect him despite correctly inferring his information.

The second statement of Theorem 2 says that any “real” competition between candidates neces-

sarily reduces the voter’s welfare relative to a non-competitive equilibrium in which one candidate

plays the unbiased strategy.18 Clearly, the latter outcome can also be achieved if there were only

one candidate contesting the election in the first place. This is a novel rationale, based on informa-

tion aggregation of politicians’ policy-relevant information, for why non-contested elections may be

beneficial to citizens.

3.3. Pandering and Welfare

A common view is that voter welfare is harmed when politicians pander to the voter’s prior (e.g.

Heidhues and Lagerlof, 2003). While we have established in Proposition 1 and Proposition 2 that

politicians may in fact be driven to anti-pander rather than pander, it is nevertheless of interest to

understand how pandering would affect voter welfare.

We find that an appropriate degree of (dis-equilibrium) pandering would actually benefit the

voter. To get some intuition for why, consider again the benchmark where both politicians play

the unbiased strategy, yi(θ) = E[θ|θi] = βα+β θi. As established in Proposition 1, the voter would

then select the politician with the more extreme platform. This implies a “winner’s curse”: the

electoral winner, say i, would have received the more extreme signal, and so voter welfare would be

improved if i were elected with a slightly more moderate platform. Such moderation can be achieved

by under-reacting to private information, i.e. pandering.

Building on this intuition, the following result shows that, subject to a qualifier, voter welfare in

the Downsian game form would be actually be maximized by a suitable degree of pandering.

18 For example, one can explicitly compute the voter’s welfare difference between a non-competitive election in whichthe winner plays the unbiased strategy and the competitive equilibrium of Proposition 2. The voter’s ex-ante expected

utility in the former is − 1α+β

, while it is − 4β2+α2+5βα(α+2β)2(α+β)

in the latter. Both expressions converge to 0 as either β →∞or α→ 0; hence, both equilibria are welfare equivalent in these limiting cases. However, the welfare difference betweenthe two equilibria is not monotonic in the parameters.

13

Proposition 3. The strategy profile where both candidates play

y(θi) = E[θ|θi, |θ−i| < |θi|], (5)

maximizes voter welfare among all candidates’ strategy profiles in which the voter’s optimal response

would lead to candidate i winning whenever |θi| > |θ−i|. This strategy displays pandering. Hence, an

appropriate degree of pandering is beneficial to voter’s welfare.

The intuition for (5) is follows: for any candidate i, the platform that maximizes voter welfare

given any information Ωi is E[θ|Ωi]; when the voter is (optimally) selecting the candidate with the

more extreme signal, the relevant information in Ωi consists of i’s own signal, θi, and what i can infer

from being selected by the voter, viz. that |θi| > |θ−i|. It is then straightforward why (5) displays

pandering: in addition to his own signal, a candidate conditions on the opponent having a more

moderate signal. Since the voter would optimally elect the candidate with a more extreme signal

if both candidates used unbiased strategies (recall Proposition 1 and its discussion), an implication

of Proposition 3 is that both candidates playing according to (5) dominates—in terms of voter

welfare—both candidates playing unbiased strategies, and hence, by Theorem 1, any equilibrium of

the office-motivated game.19

4. Discussion

In this section, we discuss three issues: more general voter preferences and information structures,

candidates who are not entirely office motivated, and alternative game forms.

4.1. Preferences and Information Structures

Our fundamental result on how electoral competition limits welfare, Theorem 1, holds rather generally

across voter preferences and information structures. An inspection of the theorem’s proof shows that

19 We conjecture that Proposition 3 also holds without the qualification that a candidate must win when he has themore extreme signal. To see some intuition, consider any symmetric strategy profile where both candidates play y(·)that is symmetric around 0. For the unbiased strategy, y(θi) = E[θ|θi], we have y′(·) = β

β+α; for the overreaction

strategy identified in Proposition 2, y(θi) = E[θ|θi, θ−i = θi], we have y′(·) = 2βα+2β

. Presuming differentiability, one

can verify that whenever y′(·) ∈ [0, 2βα+2β

], it would be optimal for the voter to elect the candidate with the moreextreme platform and hence the more extreme signal. Thus, roughly speaking, the requirement that a candidate winswhen he has the more extreme signal is satisfied so long as neither candidate overreacts by more than he would whenconditioning on the opponent having received the same signal as he did. It appears unlikely that such a degree ofoverreaction could be socially desirable.

14

the only juncture at which the information structure plays any role is to ensure that when a candidate

i’s strategy is informative, the distribution of i’s platforms from the point of view of his opponent, −i,

is not linear in θ−i. Plainly, this is a property that will typically hold in any statistical structure in

which θA and θB are correlated with each other through the true state θ. As long as this property is

satisfied, Theorem 1 holds regardless of what the statistical structure is, whether and how candidates

wish to deviate from the unbiased strategy profile, what the voter’s utility function U(y, θ) is, how

the voter randomizes if indifferent (cf. fn. 17), and what the policy space is. For example, the result

applies to the models of Heidhues and Lagerlof (2003) and Loertscher (2012), thereby generalizing

some of those authors’ observations in their models.

The results on anti-pandering do require more structure. However, they extend directly to a

well-known class of statistical structures in which the distribution of the candidates’ signals is in

an exponential family and the prior is a conjugate prior. This class includes a variety of widely-

used discrete and continuous distributions with bounded and unbounded supports, such as normal,

exponential, gamma, beta, chi-squared, binomial, Dirichlet, and Poisson. Appendix E works out a

Beta-Bernoulli example that delivers the same qualitative insights as our normal-normal structure.

The important property within the exponential family is that the posterior expectation E[θ|θ1, . . . , θn]

of the state θ given a prior mean parameter, say θ0, and any number of signal realizations, θ1, . . . , θn,

is linear in (θ0, θ1, . . . , θn) (Jewel, 1974; Kass, Dannenburg and Goovaerts, 1997). When the dis-

tribution of each θi|θ is identical for i = 1, . . . , n,20 there are constants w0 and w1 such that

E[θ|θ1, . . . , θn] = w0w0+nw1

θ0 + nw1w0+nw1

∑ni=1 θin , and therefore,

E[θ|θ1, . . . , θn]− θ0 =nw1

w0 + nw1

(∑ni=1 θin

− θ0

), (6)

while

∑ni=1 E[θ|θi]n

− θ0 =

∑ni=1

(w0

w0+w1θ0 + w1

w0+w1θi

)n

− θ0 =w1

w0 + w1

(∑ni=1 θin

− θ0

). (7)

It is immediate from (6) and (7) that for any n > 1 and any vector of signal realizations, the

average of the individual posterior expectations and the posterior expectation given the average

signal both shift in the same direction relative to the prior mean, but the latter does so by a larger

20 The points made below hold even when this is not the case, but this simplification makes the argument transparent.

15

magnitude. It is this property that underlies the incentive to overreact in an unbiased strategy profile

(Proposition 1); the logic of anti-pandering thus applies here.21,22 The following generalization of

Proposition 2 can be verified: there is an equilibrium with overreaction in which both candidates play

y(θi) = 2w1w0+2w1

θi +w0

w0+2w1θ0, with the voter being indifferent between them after observing any pair

of on-path platforms. While there may be off-path platforms in this setting (as in the Beta-Bernoulli

example in Appendix E), the equilibrium can be supported with reasonable off-path beliefs.

4.2. Ideological Motivation

Political candidates surely care about securing office. Yet, one may argue that it is extreme to

assume they are solely office motivated. Our welfare results are robust to small departures from this

assumption. Specifically, suppose that each candidate i’s payoff is

ui(y, θ,W ) = −ρi(y − θ − bi)2 + (1− ρi)1W=i, (8)

where y is the implemented policy, θ is the state, and W ∈ A,B is the winner of the election. In

this specification, ρi ∈ [0, 1] measures how policy-motivated candidate i is while bi ∈ R measures his

preference conflict with the voter over policies, i.e. his ideological bias. Assume these parameters are

common knowledge. A candidate i is office-motivated if ρi = bi = 0, he is benevolent if ρi = 1 and

bi = 0, and he is purely policy-motivated but ideologically biased if ρi = 1 and bi 6= 0. We say that

candidates whose preferences are given by (8) have mixed motivations.

An election with mixed motivations is not generally a constant-sum game for the candidates.

Thus, the analytical tools we have used earlier to characterize bounds on voter welfare cannot be

directly applied. Furthermore, because our policy space is not compact—hence standard results like

the Theorem of the Maximum cannot be invoked—it is not obvious that our welfare conclusions

remain essentially intact when candidates have even the slightest degree of policy motivation. We

formally show in the Appendix C that when bi and ρi are sufficiently close to zero for each candidate

i, any equilibrium that maximizes the voter’s welfare must have one candidate winning with ex-ante

21 Note that because the prior density need not be symmetric any longer around the mean (unlike with a normalprior) and signals may be bounded (unlike with normally distributed signals), the appropriate definitions of overreactionand pandering have to be broadened from earlier. We now say that a strategy yi(·) displays pandering if for all θi,|yi(θi) − E[θ]| ≤ |E[θ|θi] − E[θ]| with strict inequality for some θi. Analogously, yi(·) has overreaction if for all θi,|yi(θi)− E[θ]| ≥ |E[θ|θi]− E[θ]| with strict inequality for some θi.

22 The focus on posterior expectations of the state is justified when the voter has a quadratic loss function. See thediscussion in Roux and Sobel (2013) to get a sense of how asymmetric loss functions would affect the conclusions.

16

probability close to one, i.e. be almost non-competitive. Consequently, the maximum equilibrium

voter welfare must be close to that obtained by efficiently aggregating one candidate’s signal, despite

there being two sources of information. Thus, the message of Theorem 2 carries over when candidates

are largely, but not entirely, office motivated.23

4.3. Non-Binding Announcements

As discussed in the introduction, some degree of policy commitment is empirically plausible. Notwith-

standing, one may worry that our normative results, Theorem 1 and Theorem 2, rely critically on

not endowing candidates any ability to communicate their information without committing to poli-

cies. After all, if one holds fixed the information revealed through platforms (in the fully-revealing

equilibrium of Proposition 2, for example), allowing candidates to implement an arbitrary policy

once elected would yield higher welfare for the median voter.

We should note two qualifications to such reasoning. First, when the electorate consists of

heterogeneous ideologies (e.g. as formalized after Proposition 2), there will be conflict among voters

about their preferred policy, and hence it would not necessarily be straightforward for a candidate

to implement a policy different from his platform. Second, even focussing on a single (median)

voter, the voter need not be better off when the elected candidate is free to implement any policy if

candidates have their own policy preferences.24

Even setting these qualifications aside, however, and focussing on our baseline model with a single

voter and purely office-motivated candidates, we believe that relaxing the commitment assumption

would not improve voter welfare precisely because it is then implausible that office-motivated can-

23 We should emphasize that our notion of a candidate i being largely office motivated involves not only ρi ≈ 0 butalso bi ≈ 0; this plays an important role in the arguments provided in Appendix C. We cannot rule out that in agame in which ρA and ρB are both approximately zero but neither bA nor bB are, the maximum equilibrium welfarefor the voter may be “much” larger than when ρA = ρB = 0. The intuition is that competition can have a beneficial“disciplining effect” in the presence of large ideological biases, as is well understood since the work of Wittman (1977)and Calvert (1985). Nevertheless, the ideological biases would have to be large enough for this effect to outweigh thewelfare loss we identify from information distortion when candidates care significantly about winning.

24 Here is an example. For simplicity, take the policy space to be compact, say [−1, 1]. Let candidate A’s preferencesbe represented by uA(y, θ,W ) = 1W=A + y and B’s preferences be represented by uB(y, θ,W ) = 1W=B − y, whereW is the winner of the election and y is the implemented policy. Plainly, given any strategy for the voter, the candidatesare engaged in a constant-sum game. Since the constant-sum property is what underlies Theorem 1, it can be verifiedthat Theorem 1 also applies to this setting. Note that one equilibrium of this Downsian game has both candidatescommitting to the ex-ante optimal policy, 0, independent of their information. Now consider the game with non-bindingplatforms. The lack of commitment implies that the elected candidate will always implement an extreme policy (policy1 by A and policy −1 by B); it readily follows that the voter’s welfare in any equilibrium of this game may be lowerthan in (the voter-welfare-maximizing equilibrium of) the Downsian game.

17

didates reveal their private information through non-binding platforms. There are various ways in

which one can alter our model to include non-binding communication, or “cheap talk”. For con-

creteness, consider the following variation: platforms are simply non-binding, so that the elected

candidate is free to implement any policy regardless of the platform he announced. As one would

expect, this pure communication game has multiple equilibria. There are uninformative equilibria in

which the elected politician implements the best policy given (only) his information; these equilibria

provide the same welfare as the best equilibrium of our baseline model.

But there are also equilibria that fully aggregate information: each candidate can “truthfully”

reveal his information through his non-binding announcement; the voter randomizes equally between

candidates no matter their platforms; and the elected candidate implements the efficient policy

given the information revealed by both candidates’ announcements. We claim such equilibria are

not plausible, however. Intuitively, suppose that candidate i could deviate and “refuse” to reveal his

information through his non-binding announcement (or simply misrepresent his information and then

claim to have done so). Given that the opponent has revealed his information (as, by hypothesis, he

is playing his equilibrium strategy), i is strictly better informed than his opponent and is the only

candidate capable of implementing the efficient policy; the voter should thus elect i. Consequently,

either candidate would find such a deviation profitable. In fact, it is straightforward that this logic

applies not just to fully-revealing equilibria, but to any equilibrium in which either candidate makes

informative announcements. In other words, the only plausible equilibria are uninformative.

Formalizing this argument requires two ingredients: (i) ensuring it is feasible for a candidate

to deviate in the prescribed manner; and (ii) ensuring that after such a deviation, the voter will

elect the candidate who she believes to better informed. Our baseline equilibrium concept (perfect

Bayesian equilibrium) is compatible with precluding both (i) and (ii); the former because there may

be no off-path announcements, and the latter because the voter may perversely believe that the

better-informed candidate will implement a less efficient policy after the election. In Appendix D

we formalize a refinement criterion that builds on the ideas of Myerson (1989) and Farrell (1993)

to eliminate such equilibria. Indeed, in a simpler sender-receiver context, Farrell (1993, Example 2)

observed that informative equilibria are not compelling when an informed party would be better off

by not revealing any information.

Thus, we suggest, perhaps paradoxically, that it is policy commitment which allows office-

18

motivated candidates’ platforms to reveal information about their signals. The logic that a can-

didate does not want to reveal information through cheap talk because it makes his opponent better

informed—or, that by concealing his information, he can gain an informational advantage to sub-

sequently exploit—has similar ramifications for other game forms besides the pure communication

game. Consider a game where candidates can first make cheap-talk announcements followed by

binding platforms. Again, we would suggest that the only plausible equilibria are those in which no

information is revealed in the cheap-talk stage, in which case electoral outcomes coincide with those

of our Downsian model. A similar point applies if candidates are given a choice between non-binding

and binding platforms, whether or not this choice is preceded by a round (or indeed, multiple rounds)

of cheap-talk announcements.

5. Conclusion

Motivated by the debate on whether political competition promotes information aggregation and

allows an electorate to make informed choices when exercising its voting rights, this paper has

studied Downsian electoral competition between two office-motivated candidates who have private

information about policy consequences.

Contrary to a prevalent intuition that politicians are driven to policies favored by the electorate’s

prior beliefs, we find that familiar information structures impel politicians to anti-pander, i.e. to

overreact to their information. An anti-pandering equilibrium can even result in polarized policies

wherein each politician’s platform is more extreme than the preferred policy of the average voter who

votes for him. Despite revealing information, the welfare distortions associated with such platforms

can be severe.

Our main normative result is a sharp bound on voter welfare, which holds generally across infor-

mation structures: welfare in any equilibrium is effectively determined by just one candidate’s policy

function. Consequently, Downsian elections cannot efficiently aggregate more than one candidate’s

information, despite the availability of two informational sources. We also find that voter welfare is

maximized by non-competitive equilibria in which one candidate wins the election with probability

one while choosing a policy that is socially optimal based on his information alone. Any equilib-

rium in which both candidates win with positive ex-ante probability—including the anti-pandering

equilibrium—yields strictly lower voter welfare. Indeed, an appropriate degree of (dis-equilibrium)

19

pandering candidates would actually benefit voters.

We have argued that the fundamental source of inefficiency in aggregating both candidates’

information is their office motivation. Even if candidates are able to make non-binding or cheap-talk

announcements, we suggest that no information can be revealed through this channel in plausible

equilibria, and hence the maximum voter’s welfare is no larger than that of a Downsian election.

As in most formal models of spatial electoral competition, we have restricted attention to two can-

didates and assumed that their information is exogenously given. Relaxing both these assumptions

are interesting topics for future research.

Furthermore, while we have focused exclusively here on electoral competition, we suspect the

logic underlying anti-pandering and welfare limits are relevant more broadly. For example, consider

two consultants (or other experts) proposing changes that an organization should undertake; the

optimal course of action is uncertain and only one of their proposals can be accepted. One may also

consider product choice or standards submissions by firms who are vying for a consumer’s purchase

or a standards-setting body’s approval. Depending on the application, it may of interest to extend

our model to incorporate additional features such as prices.

20

Appendices

A. Two-Player Constant-Sum Games

This Appendix states a key auxiliary result, Theorem 3 below, which we apply in the proof of

Theorem 1. For finite games, the result has been obtained by Viossat (2006, Proposition 3.8); we

require a version of infinite games and provide a direct proof.

A two-player complete-information constant-sum game is given by (S, u1, u2) where S := S1×S2

with each Si a topological action (i.e. pure strategy) space for player i, and each ui : S → R is a

bounded utility function for player i such that for s ∈ S, u1(s) = −u2(s). We write ∆(Si) and ∆(S)

as the spaces of mixed strategies and mixed strategy profiles respectively,25 and payoffs are extended

to mixed strategies using expected utility as usual. For any µ ∈ ∆(S), we write µ(·|si) ∈ ∆(S−i) as

the conditional distribution of µ over S−i given si.

µ ∈ ∆(S) is a correlated equilibrium (or, more precisely, an objective correlated equilibrium)

of this game if for any i and si ∈ supp[µ],

ui(si, µ(·|si)) = sups′i∈Si

ui(s′i, µ(·|si)).

The game has a value if there exists

v∗1 := maxσ1

minσ2

u1 (σ1, σ2) = minσ2

maxσ1

u1 (σ1, σ2) .

We say that v∗1 is player 1’s value and that any solution to the above minimax/maximin problem

is an optimal strategy for player 1. Analogously, v∗2 = −v∗1 is player 2’s value and her optimal

strategies are defined in the analogous way.

Fix an arbitrary two-player complete-information constant-sum game as just defined.

Theorem 3. If µ is a correlated equilibrium, then for µ-a.e. si and s′i, ui(s′i, µ(·|si)) = v∗i (and hence

s′i is a best response to µ(·|si)).

(A proof is provided below.)

Observe that while Theorem 3 is stated for correlated equilibria of complete-information games,

it has an obvious implication for Bayesian equilibria for a class of Bayesian games. Specifically, take

any two-player constant-sum Bayesian game G ≡ (S, u1, u2, T, p) where S, u1, u2 are as introduced

before, T := T1 × T2 is the space of type profiles (each Ti is a measurable type space for player

i), and p ∈ ∆(T ) is a common-prior probability measure, so that any type’s ti’s beliefs about the

opponent’s type is given by p(t−i|ti). Note that the players’ types may be correlated. The important

restriction in G is that each player’s payoff depends only on the action profile and not on the type

profile (and that the available actions are type-independent). A mixed-strategy for player i in this

25 More precisely: each Si is viewed as a measurable space with its Borel sigma-field and ∆(Si) is the space of Borelprobability measures on Si. The product space S is endowed with the product topology.

21

Bayesian game is σi : T → ∆(S), and a Bayesian equilibrium is (σ1, σ2) that satisfies the usual

conditions. It is straightforward that any Bayesian equilibrium of G is a correlated equilibrium

of the complete-information game G ≡ (S, u1, u2). It follows from Theorem 3 that if (σ1, σ2) is a

Bayesian equilibrium of G, then for any i and for p-almost-all ti and t′i, σi(ti) must be a best response

for type t′i given the distribution p(·|t′i) over his opponent’s type and σ−i(·).

A.1. Proof of Theorem 3

The proof of Theorem 3 requires three lemmas. Throughout, we hold fixed an arbitrary two-player

complete-information constant-sum game as defined above.

Lemma 1. If µ is a correlated equilibrium, then the game has a value, and each player’s payoff from

µ is his value. Moreover, the marginal induced by µ for each player is an optimal strategy for that

player.

Proof. Let vi = ui(µ) be player i’s payoff in the correlated equilibrium; clearly v1 = −v2. Let

σi ∈ ∆(Si) be the marginal distribution over player i’s actions induced by µ. Since µ is a correlated

equilibrium, it must be that for any s1 ∈ S1, u1(s1, σ2) ≤ v1 (otherwise, for some recommendation,

s1 would be a profitable deviation), and hence u2(s1, σ2) ≥ −v1 = v2. So player 2 has a strategy

that guarantees him at least v2. By a symmetric argument, player 1 has a strategy that guarantees

him at least v1 = −v2. It follows that (v1, v2) is the value of the game; furthermore, each σi is an

optimal strategy.

Lemma 2. If µ is a correlated equilibrium, then for µ-a.e. si, ui(si, µ(·|si)) = v∗i .

Proof. By Lemma 1, the game has a value and each player has an optimal strategy. Since any

si ∈ supp[si] must be a best response against µ(·|si), it follows that

for any si ∈ supp[µ], ui(si, µ(·|si)) ≥ v∗i . (9)

But this implies that under µ, neither player can have a positive-probability set of actions that

all yield him an expected payoff strictly larger than v∗i , because then by (9) his expected payoff from

µ would be strictly larger than v∗i , implying that the opponent’s payoff from µ is strictly lower than

v∗−i, a contradiction.

For finite games, the following result was noted by Forges (1990).

Lemma 3. If µ is a correlated equilibrium, then for µ-a.e. si, µ (·|si) is an optimal strategy for

player −i.

Proof. Without loss of generality, assume i = 1. By Lemma 2, u1(s1, µ(·|s1)) = v∗1 for µ-a.e. s1.

Hence, by best responses in a correlated equilibrium, it follows that for µ-a.e. s1,

v∗1 = maxs′1

u1

(s′1, µ (·|s1)

)= −min

s′1

(−u1

(s′1, µ (·|s1)

))= −min

s′1u2

(s′1, µ (·|s1)

).

22

Since v∗1 = −v∗2, we conclude that for µ-a.e. s1, v∗2 = mins′1 u2 (s′1, µ (·|s1)) , i.e. µ(·|s1) guarantees

player 2 a payoff of v∗2, hence µ(·|s1) is an optimal strategy for player 2.

Proof of Theorem 3. Without loss of generality, let i = 1. Fix any s1 that is generic with respect

to the measure µ. From Lemma 3, we have that for any s′1, u1(s′1, µ(·|s1)) ≤ v∗1. So suppose that

there is a set, S′1, such that µ(S′1) > 0 and for each s′1 ∈ S′1, u1(s′1, µ(·|s1)) < v∗1, or equivalently that

u2 (s′1, µ(·|s1)) > v∗2. Then, there must be some set S′2 such that µ(S′2) > 0 and µ(S′1|S′2) > 0. By

playing µ(·|s1) whenever µ recommends any s2 ∈ S′2, player 2 has an expected payoff strictly larger

than v∗2 for a positive-probability set of recommendations, a contradiction with Lemma 2.

B. Proofs

Proof of Proposition 1. Assume both candidates use the unbiased strategy, i.e. yi(θi) = βα+β θi.

Since this strategy is fully revealing, the voter correctly infers θA, θB for all signal realizations. The

voter’s expected utility from a platform y given signal realizations θA and θB has the standard

mean-variance decomposition:

E[U (y, θ) |θA, θB] = −E[(y − θ)2 |θA, θB

]= −

[y2 + E(θ2|θA, θB)− 2yE(θ|θA, θB)

]= −

[y2 + E(θ|θA, θB)2 − 2yE(θ|θA, θB)

]− E(θ2|θA, θB) + E(θ|θA, θB)2

= − [y − E(θ|θA, θB)]2 − V ar (θ|θA, θB) .

This pins down the voter’s strategy; in particular, the voter must elect candidate i rather than j if

yi is closer to E[θ|θA, θB] than is yj .

We now argue that for any θA, candidate A can profitably deviate from the unbiased strategy

prescription. To show this, note that against any realization of θB, A wins if

(yB (θB)− E [θ|θB, θA])2 > (yA (θA)− E [θ|θB, θA])2 .

Substituting from the formula for the unbiased strategy and from (2), this is equivalent to(β

α+ βθB −

β

α+ 2β(θA + θB)

)2

>

(β

α+ βθA −

β

α+ 2β(θA + θB)

)2

,

or after algebraic simplification, (θA)2 > (θB)2. Hence, A wins when his type is more extreme (i.e.

larger in magnitude) than B’s, which implies that no matter his true type, candidate A strictly

increases his win probability by mimicking a more extreme type.

We require two lemmas to prove our main results. Lemma 4 below shows that a candidate’s

probability of winning with any on-path platform cannot depend on his signal. Some notation is

helpful to state the result precisely. Given any equilibrium (which may have mixing by candidates),

let Πi(yi; θi) denote the expected utility (i.e. win probability) for candidate i when his type is θi and

he plays platform yi, and let Yi denote the set of platforms that i plays with strictly positive ex-ante

23

probability. Given any equilibrium, when we refer to “for almost all on-path platforms”, we mean

for all but a set of platforms that have ex-ante probability zero with respect to the prior over types

and the equilibrium strategies. Similarly for statements about generic platforms.

Lemma 4. Given any equilibrium and any i, for almost all on-path platforms, yi, y′i, and almost all

types θi, θ′i, Πi (yi; θi) = Πi (y′i; θ

′i) .

Proof. Fix any equilibrium. Given the voter’s strategy, p(yA, yB), the induced game between the two

candidates is a constant-sum Bayesian game. Any equilibrium of this two-player Bayesian game is

a correlated equilibrium of a complete-information constant-sum game between the two candidates

where each chooses an action yi ∈ R and for any profile (yA, yB), the payoff to candidate A is

p(yA, yB) while the payoff to candidate B is 1 − p(yA, yB). The Lemma follows from a general fact

about constant-sum games that is stated and proved as Theorem 3 in Appendix A.

We can now show that any informative equilibrium is an ex-post equilibrium in the sense that

neither candidate can affect his probability of winning by changing his platform, no matter which of

his opponent’s platforms is realized.

Lemma 5. In any informative equilibrium, the voter’s strategy is constant over almost all on-path

platforms.

Proof. Fix any equilibrium where, without loss of generality, candidate B is playing an informative

strategy. We will show that for a generic platform of candidate A, A’s winning probability is almost-

everywhere constant over B’s platforms. Pick an arbitrary finite partition of the range of player B’s

on-path platforms, Y 1B, . . . , Y

mB , where each Y j

B is a convex set. Without loss, we may take m > 1

since B’s strategy is informative.

For a generic on-path platform of player A, yA, Lemma 4 implies there is some v∗A such that

v∗A = ΠA (yA; θA) = ΠA

(yA; θ′A

)for almost all θA, θ

′A.

Let q(Y jB|θA) be the probability that B plays a platform in the set Y j

B given his possibly-mixed

strategy σB(·) and that his type is distributed according to the conditional distribution given θA, i.e.

θB|θA ∼ N(

βα+β θA,

βα+β + 1

β

). Let p(yA|Y j

B; θA) be the probability with which A of type θA expects

to win when he chooses platform yA given that the opponent’s platform falls in the set YB; notice

that because p(·, ·) is locally constant (as our restriction is to equilibria where p(yA, ·) has only at

most a countable number of discontinuities), the dependence on θA can be dropped if each Y jB has

been chosen as a sufficiently small interval, because then the distribution of B’s platforms within Y jB

is irrelevant. Therefore, with the understanding that each Y jB is a small enough interval, we write

p(yA|Y jB).

Therefore, for any generic m types of player A, (θ1A, . . . , θ

mA ), we have q

(Y 1B|θ1

A

)· · · q

(Y mB |θ1

A

)...

...

q(Y 1B|θmA

)· · · q (Y m

B |θmA )

p

(yA|Y 1

B

)...

p (yA|Y mB )

=

v∗A...

v∗A

.

24

The unknowns above are p(yA|Y jB); clearly one solution is for each p(yA|Y j

B) = v∗A. If we prove

that this is the unique solution for some generic choice of (θ1A, . . . , θ

mA ), we are done, because yA was a

generic platform for A and the partition Y 1B, . . . , Y

mB was arbitrary (subject to each Y j

B being a small

enough interval). The Rouche-Capelli Theorem implies that it suffices to show that for some choice of

(θ1A, . . . , θ

mA ), the coefficient matrix of q(·|·) above has non-zero determinant. Suppose that for some

selection of distinct types (θ1A, . . . , θ

mA ), the coefficient matrix has zero determinant. Since q(·|θjA)

changes non-linearly in θjA (because B’s strategy is informative and θB|θjA is normally distributed),

it follows that the determinant cannot remain zero for all perturbations of (θ1A, . . . , θ

mA ).

We next prove a result that will deliver Proposition 2 as a corollary. To state the result, we

need some terminology. Say that a (possibly mixed) strategy for candidate i, σi : R → ∆(R), is

locally pure at θi if types in a neighborhood of θi do not mix, and hence it is meaningful to write

yi(·) in a neighborhood of θi. A strategy that is locally pure at θi is locally continuous at θi if

yi(·) is continuous in some neighborhood of θi. A strategy that is locally pure at θi is also locally

revealing at θi if there there is a neighborhood of θi, say N(θi), such that for each θ′i ∈ N(θi),

θi ∈ R : yi(θ′i) ∈ supp[σi(θi)] = θ′i. In words, local revelation requires that the voter be able

to infer a candidate’s signal from his platform for some interval of signals; it is then well-defined to

write the inverse function y−1i (yi(θi)) in the relevant region.

Proposition 4. Any competitive equilibrium in which for each candidate i ∈ A,B there is some

θi such that i’s strategy is locally pure, revealing, and continuous at θi must be such that for some

i ∈ A,B and c ∈ R:

yi(θi) =2β

α+ 2βθi + c and y−i(θ−i) =

2β

α+ 2βθ−i − c. (10)

Conversely, for any i ∈ A,B and c ∈ R, the above strategy profile defines an equilibrium where the

voter elects each candidate with probability 1/2 after observing any platform pair (yA, yB) ∈ R2.

Proof. We prove the second statement first. Given the voter’s behavior, a candidate’s platform does

not affect his election probability. It suffices therefore to verify that the voter is indifferent between

the two candidates for any pair of platforms. Since the candidates’ strategies are fully revealing,

the voter correctly infers the candidates’ signals from the platform pair. Furthermore, since the

candidates’ pure strategies each have range R, there are no off-path platform pairs. Therefore,

it suffices to show that for any θi and θ−i, −E[(yi(θi) − θ)2|θi, θ−i] = −E[(y−i(θ−i) − θ)2|θi, θ−i],or equivalently that (yi (θi)− E [θ|θi, θ−i])2 = (y−i (θ−i)− E [θ|θi, θ−i])2.26 Using (10) and (2), this

latter equality can be rewritten as(2β

α+ 2βθi + c− 2β

α+ 2β

(θi + θ−i

2

))2

=

(2β

α+ 2βθ−i − c−

2β

α+ 2β

(θi + θ−i

2

))2

,

which holds for any θi, θ−i.

26 That this latter equality is equivalent to the former follows from a standard mean-variance decomposition underquadratic loss utility as in the proof of Proposition 1.

25

To prove the first part of the result, fix a competitive equilibrium such that each candidate

i’s strategy is locally continuous, revealing, and pure in some neighborhood of some point, θi; it

is then well-defined to write yi(·) in the neighborhood. It also follows that for each i there is a

non-empty, open, and convex set Yi ⊆ range[yi(·)] such that the set of platforms in Yi has positive

probability under yi(·) and yi(·) is invertible over Yi. Hence, for any yi ∈ Yi, θi(yi) := y−1i (yi) is

well defined. Since the equilibrium is informative (by local revelation) and competitive, Lemma 5

implies that the voter randomizes equally between both candidates for all on-path platform pairs.

This implies that for any y′A ∈ YA and y′B ∈ YB, we must have E[θ|y′A, y′B] = y′A+y′B2 , which implies

βα+2β (θA(y′A) + θB(y′B)) = y′A+y′B

2 , or equivalently

θB(y′B) =α+ 2β

2β

(y′A + y′B

)− θA(y′A). (11)

Now observe that since each Yi is an open set, given any yA ∈ YA and yB ∈ YB, the same argument

can also be made for platforms yA + ε and yB − ε for all ε that are small enough in absolute value.

Hence,

θB(yB − ε) =α+ 2β

2β(yA + yB)− θA(yA + ε). (12)

Substituting into (12) from (11) with y′B = yB − ε and y′A = yA yields

α+ 2β

2β(yA + yB − ε)− θA(yA) =

α+ 2β

2β(yA + yB)− θA(yA + ε),

or equivalently,

θA(yA + ε) =α+ 2β

2βε+ θA(yA). (13)

Since yA and ε were arbitrary (so long as yA ∈ YA and |ε| is small), the equality in (13) re-

quires that for some constant cA ∈ R, θA(yA) = α+2β2β yA + cA for all yA ∈ YA, or equivalently

yA(θA) = 2βα+2β θA + cA. A symmetric argument establishes the analog for yB(θB) with some con-

stant cB. But now (11) requires cB = −cA. Finally, the two strategies must be pure and linear

on the entire domain, because otherwise, given the linearity on a subset of the domain, there will

be a positive-probability set of on-path platform pairs, say (yA, yB), for which E[θ|yA, yB] 6= yA+yB2 ,

contradicting the voter randomizing between candidates for such platform pairs.

Proof of Proposition 2. Observe that each candidate wins with ex-ante probability 1/2 in any

symmetric equilibrium, hence a symmetric equilibrium is competitive. Both the existence and unique-

ness claims now follow as a corollary of Proposition 4.

Proof of Corollary 1. The first statement follows from the observations that the candidates’

strategies make the median voter (with ideology zero) indifferent between the two platforms no

matter their realization (Proposition 2) and that the ordering of voters’ by their interim preferred

policies is the same as by their ideologies.

For the second statement, consider any pair of realized policies, yA and yB. Given the candidates’

strategies, E[θ|yA, yB] = yA+yB2 . Thus, the interim preferred policy for voter k is yk = E [θ|yA, yB] +

bk = yA+yB2 + bk. Suppose first that i is the candidate who is further to the right: yi > y−i. The

26

interim preferred policy of the average voter who votes for i is E [yk|bk > 0] = yi+y−i2 + E [bk|bk > 0].

The difference between candidate i’s platform and his average voter’s interim preferred policy is

yi − E [yk|bk > 0] = yi−y−i2 − E [bk|bk > 0]. Thus, i’s policy is to the right of his average voter’s

interim preferred policy if and only if

yi − y−i2

> E [bk|bk > 0] . (14)

Following an analogous reasoning, if yi < y−i then i’s platform is to the left of his average voter’s

interim preferred policy if and only if

yi − y−i2

< E [bk|bk < 0] . (15)

Due to the symmetry of F around 0, inequalities (14) and (15) can be jointly expressed as|yi−y−i|

2 > E[bk|bk > 0].

Proof of Theorem 1. Fix any equilibrium. The result obviously holds if the voter is electing one

candidate with probability one on the equilibrium path, so assume this is not the case. There are

two cases to consider:

(i) First suppose the equilibrium is informative. Then Lemma 5 implies that the voter is indif-

ferent between the two candidates for almost all on-path platform pairs. Hence, voter welfare would

not change if, holding fixed the candidates’ strategies, either of the candidates were elected with

probability one.

(ii) Now suppose the equilibrium is uninformative. Then the voter chooses between platforms

based on the prior. Let V (y) denote the expected utility for the voter from policy y based on the

prior. Since the equilibrium is competitive, there exists some V ∗ such that for almost all policies y

that are proposed in equilibrium by either candidate, V (y) = V ∗; otherwise, one of the candidates

is not playing a best response to the other. This implies that the voter’s welfare would not change

if, holding fixed the candidates’ strategies, either candidate were elected with probability one.

Proof of Theorem 2. Suppose candidate i plays the unbiased strategy, yi(θi) = βα+β θi, while

candidate −i plays y−i(θi) = θi. Given these strategies, it is straightforward to verify that it is

optimal for the voter to elect candidate i no matter which pair of platforms is observed. In fact,

letting i = A without loss of generality,

(yA − E [θ|yA, yB])2 − (yB − E [θ|yA, yB])2

=

(β

α+ βθA −

β (θA + θB)

α+ 2β

)2

−(θB −

β (θA + θB)

α+ 2β

)2

≤ 0.

In turn, given this voter strategy, each candidate is obviously indifferent over all platforms, and hence

this constitutes a fully-revealing non-competitive equilibrium in which the winning candidate i plays

the unbiased strategy.

Theorem 1 implies that no equilibrium can provide the voter a strictly higher welfare than that

27

obtained in the above equilibrium; clearly any uninformative equilibrium provides strictly lower

welfare. Now consider any informative and competitive equilibrium. By Lemma 5, the voter must

be indifferent between almost-all on-path platforms, and hence the voter’s welfare can be evaluated

by assuming that the voter elects either of the two candidates with probability one (while holding

their strategies fixed). But then, since both candidates cannot be playing the unbiased strategy

(Proposition 1), it follows that the voter’s welfare is strictly lower than in the above equilibrium.

Proof of Proposition 3. By the law of iterated expectations, the voter’s ex-ante utility can be

expressed as

E[U ] = −E[(y − θ)2

]= −E

[E[(y − θ)2 |θA, θB

]]= −E

[(y − β (θA + θB)

α+ 2β

)2]− 1

α+ 2β

= −Pr (A wins)E

[(yA −

β (θA + θB)

α+ 2β

)2 ∣∣∣ A wins

]

− Pr (B wins)E

[(yB −

β (θA + θB)

α+ 2β

)2 ∣∣∣ B wins

]− 1

α+ 2β. (16)

It is convenient to define hi (θi) := E [θ−i|θi, i wins]. Using iterated expectations again and a

mean-variance decomposition as in the proof of Proposition 1, it also holds that for any i ∈ A,B,

E

[(yi −

β (θA + θB)

α+ 2β

)2 ∣∣∣ i wins]

= E

[E

[(yi −

β (θA + θB)

α+ 2β

)2 ∣∣∣ θi, i wins] ∣∣∣ i wins]

= E

[(yi −

β (θi + E [θ−i|θi, i wins])α+ 2β

)2

+

(β

α+ 2β

)2

V ar [θ−i|θi, i wins]∣∣∣ i wins]

= E

[(yi −

β (θi + h(θi))

α+ 2β

)2 ∣∣∣ i wins]+

(β

α+ 2β

)2

E[V ar [θ−i|θi, i wins]

∣∣∣ i wins] . (17)

(16) and (17) imply

E[U ] = −(

β

α+ 2β

)2

LV − LE −1

α+ 2β, (18)

where

LV :=∑i=A,B

Pr (i wins)E[V ar [θ−i|θi, i wins]

∣∣∣ i wins] , (19)

LE :=∑i=A,B

Pr (i wins)E

[(yi(θi)−

β (θi + h(θi))

α+ 2β

)2 ∣∣∣ i wins] . (20)

Our problem is to maximize (18) subject to i winning when |θi| > |θ−i|. Since (19) does not

28

depend on platforms while (20) is bounded below by 0, a solution must satisfy for each i:

yi(θi) =β (θi + h(θi))

α+ 2β= E[θ|θi, i wins].

Since the constraint is that i wins when |θi| > |θ−i|, it follows immediately that the solution is for

each candidate to use the strategy (5).

Using the closed-form expression for truncated normal distributions, equation (5) can be expressed

as

y (θi) =β

α+ βθi − σ

β

α+ 2β

φ(

1σ

αα+β θi

)− φ

(− 1σα+2βα+β θi

)Φ(

1σ

αα+β θi

)− Φ

(− 1σα+2βα+β θi

) ,where σ =

√α+2β

(α+β)β , and φ(·) and Φ(·) are respectively the density and cumulative distributions of

the standard normal distribution, N (0, 1). To see that this strategy has pandering, consider any

θi > 0 (with a symmetric argument for θi < 0). Then 0 < y (θi) <β

α+β θi because φ(

1σ

αα+β θi

)>

φ(

1σα+2βα+β θi

)> 0 and Φ

(1σ

αα+β θi

)> Φ

(− 1σα+2βα+β θi

)> 0.

C. Mixed Motives

The goal of this section is to substantiate the discussion in Subsection 4.2 of the main text by

formally generalizing our main welfare conclusions to a setting where candidates are largely but not

entirely office motivated. The main result, Theorem 4 below, will be that when bi and ρi (as defined

in Subsection 4.2) are sufficiently close to zero for each i = A,B, any equilibrium that maximizes

the voter’s welfare must have one candidate winning with ex-ante probability close to one, i.e. be

almost non-competitive. Consequently, the maximum equilibrium voter welfare must be close to that

obtained by efficiently aggregating only one candidate’s signal.

In the context of mixed-motivation games, as defined in Subsection 4.2 of the main text, we say

that candidate’s i strategy is unbiased if

yi(θi) =β

α+ βθi + bi. (21)

Note that this refers to candidate i choosing a policy that maximizes his preference over policy given

his signal, as opposed to the voter’s.

Proposition 5. Assume candidates have mixed motivations. There is a fully revealing non-competitive

equilibrium in which one candidate i plays the unbiased strategy (21), the other candidate −i plays

y−i(θ−i) = θ−i −α+ β

βbi, (22)

and the voter elects candidate i no matter the pair of platforms. Furthermore, this is the unique fully

revealing and non-competitive equilibrium in which the winning candidate plays the unbiased strategy.

(The proof is in the Supplementary Appendix.)

29

As the equilibrium constructed above is invariant to ρA and ρB, it has a number of interesting

implications. First, the equilibrium exists when candidates are purely policy-motivated. Second,

for ρA = ρB = bA = bB = 0, this equilibrium reduces to one that verifies the first statement of

Theorem 2.27 Moreover, by taking bA = bB = 0 and ρA = ρB = 1, we see that there is also a

non-competitive equilibrium where the winner plays the unbiased strategy when both candidates are

benevolent.28 Hence, the equilibrium of Proposition 5 continuously spans all three polar cases of

candidate motivation.

We are now ready to establish the robustness of Theorem 2 when bi and ρi are sufficiently close

to zero for both i = A,B. Let an arbitrary Downsian game with mixed-motivated candidates be

parameterized by (ρ, b), where ρ = (ρA, ρB) and b = (bA, bB). Given any equilibrium, σ, of a mixed-

motivations game (including equilibria where candidates play mixed strategies), let UV (σ) be the

voter’s welfare, i.e. ex-ante expected utility, in this equilibrium. Note that the voter’s welfare depends

only on the strategies used and not directly on the candidates’ motivations. For any ε ∈ [0, 1/2], let

Σε(ρ, b) be the set of equilibria in a mixed-motivations game where each candidate wins with ex-ante

probability at least ε. Let U εV (ρ, b) := supσ∈Σε(ρ,b)

UV (σ) be the highest welfare for the voter across

all equilibria in which each candidate wins with probability at least ε ∈ [0, 1/2], given candidates’

motivations (ρ, b); in particular, U0V (ρ, b) is the highest welfare across all equilibria.

Theorem 4. Assume candidates have mixed motivations. Then,

1. As (ρ, b)→ (0,0), U0V (ρ, b)→ U0

V (0,0).

2. For any ε > 0, there exists δ > 0 such that for all (ρ, b) close enough to (0,0),29 it holds that

U εV (ρ, b) < U0V (ρ, b)− δ.

(The proof is in the Supplementary Appendix.)

To interpret this result, let σB(ρ, b) be the equilibrium identified in Proposition 5 where, without

loss, A is winning candidate who plays (21). We know from Theorem 2 that U0V (0,0) = UV (σB(0,0)).

Since Proposition 5 assures that σB(ρ, b) ∈ Σ0(ρ, b), the first part of Theorem 4 implies that when

candidates are close to purely office-motivated, the equilibrium of Proposition 5 provides close to the

highest possible equilibrium welfare to the voter. In this sense, the conclusion of Theorem 2 that the

voter’s welfare cannot be higher than in a non-competitive equilibrium in which one candidate plays

the unbiased strategy is robust to candidates having mixed motivations. It is worth remarking that

one cannot just invoke the Theorem of the Maximum here; in fact, because the policy space is not

compact, the equilibrium correspondence is not upper semi-continuous.30 However, the proof shows

27 Proposition 5 shows that there are in fact a continuum of fully revealing non-competitive equilibria with purelyoffice-motivated candidates, because when ρA = ρB = 0, one may substitute any constant in place of bi in (21) and(22) and produce a non-competitive equilibrium. However, among these, only the equilibrium in which the constant iszero has the winning candidate playing an unbiased strategy when bA = bB = 0.

28 Of course, by Proposition 3, this is Pareto-dominated when the candidates are benevolent.

29 I.e. if for each i, ρi ∈ [0, 1] and bi ∈ R are both close enough to zero.

30 To see this, note that given any candidates’ motivations with bA > 0, there is an equilibrium where both candidatesuse the constant strategy y(·) = 1/bA; this is supported by suitable off-path beliefs such that any candidate whoseplatform differs from bA loses for sure. As bA ↓ 0, this sequence of equilibria does not converge to a limit equilibrium.

30

that any sequence of voter-welfare-maximizing equilibria must converge in welfare.

The second part of Theorem 4 shows that the second part of Theorem 2 is also robust. It says

that given any ε > 0, there is a bound δ > 0 such that any equilibrium in which both candidates win

with ex-ante probability greater than ε when they are largely office-motivated will provide a level of

voter welfare that is bounded away from U0V (0,0) by δ. Combined with the first part of Theorem 4,

it follows that once candidates are primarily office-motivated, any equilibrium that maximizes the

voter’s welfare must have one candidate winning with ex-ante probability close to one and hence be

almost non-competitive.

D. A Refinement for the Communication Game

This Appendix formalizes the claim made in Subsection 4.3 that one can build on the ideas of Myerson

(1989) and Farrell (1993) to provide a formal equilibrium refinement eliminating informative equilib-

ria when candidates’ announcements are non-binding or cheap talk. We develop the formalization for

the main variation discussed in Subsection 4.3, in which candidates make (one round of) non-binding

announcements, and the elected candidate is free to choose any policy. However, it bears emphasis

that related formalizations are also available from the authors on request for other variations of the

game form, for example when candidates first make observable non-binding announcements followed

by commitments before the voter chooses whom to elect.

The communication game form we study is as follows: after privately observing their signals

θi and θ−i as in the baseline model, both candidates simultaneously choose non-binding platforms,

which we refer to hereafter as announcements, yi ∈ R and y−i ∈ R. (Considering other spaces of

announcements would not affect our results.) Upon observing the pair of announcements, the voter

updates her belief about θ and then elects one of the two candidates. As in the baseline model, the

voter elects each candidate with equal probability when she is indifferent. The elected candidate i

can choose any policy in R to implement.

Consider the equilibrium that fully aggregates information, as described in Subsection 4.3. Each

candidate i adopts the “truthful” announcement strategy, yi(θi) = θi; the voter elects each candidate

with equal probability no matter what announcements are made; and, for all (θi, θ−i), the elected

candidate i implements policy using the function si(θi, yi, y−i) = E[θ|θi, θ−i = y−i]. Notice that this

profile of strategies forms an equilibrium because neither candidate i can affect his probability of

being elected by changing his announcement.

However, we believe this equilibrium is not plausible because each candidate i would profitably

deviate by “refusing” to reveal his information, if he only somehow could, and by then implementing

the efficient policy E[θ|θi, θ−i = y−i] if elected. If the voter knows that candidate i has deviated

in this fashion, the voter would find it uniquely optimal to elect i. Note that it is credible for i to

implement policy E[θ|θi, θ−i = y−i] when elected because i is indifferent over policies once in office.

Furthermore, given that −i’s announcement has revealed θ−i but θi remains private information to i

(owing to his “refusal” to reveal his information), only candidate i can implement the efficient policy.

As discussed in the main text, there are two issues in formalizing this reasoning. To address

31

them, we appeal to the notion of neologism proofness by Farrell (1993); sea also Myerson’s (1989)

coherent plans and Matthews, Okuno-Fujiwara and Postlewaite’s (1991) announcement-proofness.

Farrell (1993) refinement is defined for cheap-talk games with one sender and one receiver. But it

can be be adapted to the current two-sender context by enriching it with Myerson’s (1989) notion

of reliable promises, as follows.

Definition of neologism-proof equilibrium. For each player i and (measurable) set of types

Θi ⊆ R, a neologism is a message, denoted as ni(Θi), while a promise is a policy strategy qi(·)mapping each type θi ∈ Θi and each pair of announcements (yi, y−i) into a policy in R. A statement

is a neologism-promise pair. Fix any equilibrium of the communication game. Assume that for each

candidate i and any set Θi, there is an out-of-equilibrium message ni(Θi). Also assume that if such a

neologism were sent by i, together with a promise qi(·), the voter would believe them in the following

sense. First, she would update her beliefs about the state θ based solely on the event θi ∈ Θi and y−i(using the equilibrium strategy of −i); furthermore, the voter would posit that candidate −i updates

analogously. Second, she would believe that player i plays according to the promise qi(·) if elected.

The promise qi(·) is reliable if it is optimal for candidate i to follow when elected. In our setting,

it is trivially true that any promise qi(·) is reliable, simply because candidates do not have policy

preferences. The statement (ni(Θi), qi(·)) is credible for candidate i (relative to the equilibrium),

if two conditions are met: (i) the promise qi(·) is reliable; and (ii) when the voter believes the

statement in the sense defined above, the set of signals θi for which candidate i strictly prefers to

send the statement (ni(Θi), qi) over his equilibrium announcement yi(θi) is precisely Θi. To be clear:

we mean “strictly prefer” in terms of candidate i’s interim expected utility given his signal, taking

expectations over candidate −i’s behavior.31 An equilibrium is neologism proof if neither candidate

has a credible statement.

Implications. We claim the “truthful” equilibrium described earlier is not neologism proof. Sup-

pose that either candidate i sends the statement composed of the neologism ni(R) and the promise

qi(·) such that qi(θi, yi, y−i) = E[θ|θi, θ−i = y−i], for all θi, yi, y−i. The neologism ni(R) can be inter-

preted as “silence”: candidate i refuses to reveal any information by saying only that his signal lies

in R. With the promise qi(·), the candidate i conveys to the voter that, if he is elected, he will use his

information θi and the opponent’s information θ−i (inferred from the announcement y−i) optimally

for voter welfare.

Plainly, the uniquely optimal action for the voter when this statement is received, and believed

in the sense defined above, is to elect candidate i for any candidate −i’s equilibrium announcement

y−i: candidate i’s post-electoral policy will be the policy that efficiently aggregates all information,

E[θ|θi, θ−i]; while −i’s policy will at best reflect only his own information, E[θ|θ−i]. Consequently, no

matter his signal θi, candidate i strictly prefers sending the statement (n(R), qi(·)) to his equilibrium

announcement. It follows that the statement (n(R), qi) is credible, which implies the equilibrium is

31 But, in fact, our conclusions would not change if one instead uses the following more demanding notion: given hissignal θi, candidate i strictly prefers sending the neologism/promise pair if (i) his payoff is weakly higher no matterwhich equilibrium announcement is sent by candidate −i, and (ii) strictly higher for a set of announcements that havepositive probability given θi.

32

not neologism proof.

The same logic extends to all fully- and partially-informative equilibria, as summarized below.

Note that the voter’s welfare in any uninformative equilibrium can be no higher than efficiently

aggregating one candidate’s signal.

Theorem 5. No informative equilibrium of the communication game is neologism proof. There exist

neologism-proof uninformative equilibria in which voter welfare is that of efficiently aggregating one

candidate’s signal.

Proof of Theorem 5. Step 1: We first show that no informative equilibrium is neologism proof.

Lemma 5 applied to the communication game implies that in any informative equilibrium, each

candidate wins with either probability one half, zero or one, no matter the candidates’ signals and

(on-path) announcements. Consider any such informative equilibrium. At least one of the two

candidates, say i, wins the game with probability at most one half regardless of his signal.

Consider first the case in which −i plays an informative strategy. Suppose i adopts the statement

composed of the neologism n(R) and the promise qi(·) such that qi(θi, yi, y−i) = E[θ|θi, y−i], for all

θi, yi, y−i. The voter then believes that if −i is elected, he will at best implement policy E[θ|θ−i];whereas if i is elected, he will implement E[θ|θi, y−i]. Following a similar decomposition to that used

in the proof of Proposition 3, the voter’s expected payoff from electing i is therefore

−E[(E [θ|θi, y−i]− θ)2 |y−i

]= − 1

α+ 2β−(

β

α+ β

)2

E[Var [θ−i|θi, y−i]

∣∣∣y−i] , (23)

Similarly, the expected payoff from electing −i is

−E[(E [θ|θ−i]− θ)2 |y−i

]= −E

[(E [θ|θ−i]− θ)2

]= − 1

α+ 2β−(

β

α+ β

)2

E [Var [θi|θ−i]]

= − 1

α+ 2β−(

β

α+ β

)2

E [Var [θ−i|θi]] , (24)

where the first equality is because y−i does not affect either expectation on the left hand side, and

the third equality just changes variables. It is clear that (23) must be weakly larger than (24) no

matter the value of y−i since the voter’s payoff can never be reduced by observing an additional

signal about θ. We now argue that because candidate −i’s strategy is informative, (23) must be

strictly larger than (24) for some on-path y−i.

Suppose this is not the case. Then, considering only on-path y−i’s,

Eθi[Var [θ−i|θi, y−i]

∣∣∣y−i] = Eθi [Var [θ−i|θi]] for all y−i

=⇒ Ey−iEθi[Var [θ−i|θi, y−i]

∣∣∣y−i] = Eθi [Var [θ−i|θi]]

⇐⇒ EθiEy−i[Var [θ−i|θi, y−i]

∣∣∣y−i] = Eθi [Var [θ−i|θi]] by changing order of integration

=⇒ Var [E [θ−i|θi, y−i]] = 0 for all θi by law of total variance

⇐⇒ E[θ−i|θi, y−i] = E[θ−i|θi] for all y−i and θi by iterated expectation,

33

which cannot hold since y−i is the realization of an informative signal about θ−i. Specifically, consider

the case where candidate −i plays a pure strategy, y−i(·). Then, since y−i(·) is not constant (because

it is informative),

E[θ−i|θi, y−i] =

∫y−1−i (y−i)

θ−if(θ−i|θi)Pr(y−i)

dθ−i 6=∫ ∞−∞

θ−if(θ−i|θi)dθ−i = E[θ−i|θi],

for a set of y−i that occur with positive probability.

Therefore, we conclude that following the statement (n(R), qi(·)) from candidate i, electing i

is at least as good for the voter as electing −i for any y−i, and strictly so for a set of y−i with

positive probability. Since candidate i wins with probability at most one half no matter y−i on the

equilibrium path, it follows that sending the statement (n(R), qi(·)) improves i’s payoff regardless of

his signal θi. Hence, the statement (n(R), qi(·)) is credible for i, which implies the equilibrium is not

neologism proof.

Consider now the case in which −i plays an uninformative announcement strategy. Then i’s

announcement must be informative (since the equilibrium is informative). If −i wins with probability

one, then it is easily verified that it is credible for i to make the same statement as above, (n(Θi), qi(·)).If −i wins with probability one half or zero, the same argument as above can be applied—reversing

the role of −i and i—to conclude that −i has a credible statement.

Step 2: Now we prove the theorem’s second statement. Consider the uninformative competitive

equilibrium in which both candidates i adopt the platform E[θ|θi] if elected, no matter the announce-

ments. It can be verified using the same logic as above that if either candidate i were to send any

neologism ni(Θi), together with any promise qi(·), the voter would weakly prefer to elect candidate

−i no matter which announcement y−i is observed. As each candidate wins with probability one half

no matter their announcements on the equilibrium path, there is no credible statement, and so the

equilibrium is neologism proof. Plainly, the voter’s welfare in this equilibrium is that of efficiently

aggregating one candidate’s signal.

E. A Beta-Bernoulli Example

We provide here an example when the state follows a Beta distribution and each candidate gets a

binary signal drawn from a Bernoulli distribution; the feasible set of policies is [0, 1] (or any superset

thereof). This statistical structure is a member of the exponential family with conjugate priors

discussed in the main text. Aside from illustrating how the incentives to overreact exist even when

the state distribution may not be unimodal and may be skewed, signals are discrete, etc., it also

provides a closer comparison with the setting of Heidhues and Lagerlof (2003) and Loertscher (2012)

than does our baseline normal-normal model.

Assume the prior distribution of θ is Be(α, β), which is the Beta distribution with parameters

α, β > 0 whose density is given by f (θ) = θα−1(1−θ)β−1

B(α,β) , where B(·, ·) is the Beta function.32 Thus

θ has support [0, 1] and E[θ] = αα+β . For reasons explained at the end of the section, we assume

32 If α and β are positive integers then B (α, β) = (α−1)!(β−1)!(α+β−1)!

.

34

α 6= β. (This rules out a uniform prior, which corresponds to α = β = 1.) Each candidate i ∈ A,Bobserves a private signal θi ∈ 0, 1; conditional on θ, signals are drawn independently from the same

Bernoulli distribution with Pr(θi = 1|θ) = θ. The policy space is any subset of R containing [0, 1].

It is well-known that the posterior distribution of the state given signal 1 is now Be (α+ 1, β)

(i.e. has density f (θ|θi = 1) = θα(1−θ)β−1

B(α+1,β) ); similarly the posterior given signal 0 is Be (α, β + 1). It

is also straightforward to check that the posterior distribution of the state given two signals is as

follows: if both θi = θ−i = 1, it is Be (α+ 2, β); if θi = 0 and θ−i = 1, it is Be (α+ 1, β + 1); and if

θi = θ−i = 0, it is Be (α, β + 2) .

It follows that

E [θ|θi] =α+ θi

α+ β + 1and E [θ|θi, θ−i] =

α+ θi + θ−iα+ β + 2

.

The above formulae imply that for any realization (θA, θB),

sign (E [θ|θA, θB]− E [θ]) = sign

(E [θ|θA] + E [θ|θB]

2− E [θ]

),

|E [θ|θA, θB]− E [θ]| >∣∣∣∣E [θ|θA] + E [θ|θB]

2− E [θ]

∣∣∣∣ . (25)

In other words, both the posterior mean given two signals and the average of the individual posterior

means shift in the same direction from the prior mean, but the former does so by a larger amount.

Consequently, if candidates were to play unbiased strategies and the voter best responds, then

whenever θA 6= θB there is one candidate who wins with probability one: the candidate i with

θi = 1 (resp., θi = 0) when β > α (resp., β < α). Of course, when θA = θB, both candidates

would choose the same platform and win with equal probability. It is worth highlighting that

when θA 6= θB, it is the candidate with the ex-ante less likely signal who wins, because ex-ante

Pr(θi = 1) = E[θ] = α/(α+β). This implies that unbiased strategies cannot form an equilibrium, but

not because candidates would deviate when drawing the ex-ante less likely signal; rather, they would

deviate when drawing the ex-ante more likely signal to the platform corresponding to the ex-ante

less likely signal.33 Notice that this profitable deviation given signal θi is to an (on-path) platform yisuch that |yi −E[θ]| > |E[θ|θi]−E[θ]|; hence, it is a profitable deviation through overreaction rather

than pandering.

Finally, we observe there is symmetric fully revealing equilibrium with overreaction in which both

candidates play

y(1) =α+ 2

α+ β + 2and y(0) =

α

α+ β + 2.

This strategy displays overreaction because

y(0) < E[θ|θi = 0] < E[θ] < E[θ|θi = 1] < y(1).

33 See Che, Dessein and Kartik (2013) for an analog where options that are “unconditionally better-looking” neednot be “conditionally better-looking”.

35

It is readily verified that when both candidates use this strategy, E [θ|θA, θB] = y(θA)+y(θB)2 for all

(θA, θB), and hence each candidate would win with probability 1/2 for all on-path platform pairs; a

variety of off-path beliefs can be used to support the equilibrium.

Note that this overreaction equilibrium would exist even when α = β. However, were α = β,

unbiased strategies would also constitute an equilibrium: the reason is that in this case, both sides of

(25) would be equal to each other (in fact, equal to zero) when θA 6= θB, and hence the voter would

elect both candidates with equal probability no matters their platforms under unbiased strategies.

36

References

Acemoglu, Daron, Georgy Egorov, and Konstantin Sonin, “A Political Theory of Populism,”

Quarterly Journal of Economics, 2013, 128 (2), 771–805.

Alesina, Alberto, “Credibility and Policy Convergence in a Two-Party System with Rational

Voters,” American Economic Review, September 1988, 78 (4), 796–805.

and Richard Holden, “Ambiguity and Extremism in Elections,” June 2008. unpublished.

Althaus, Scott L., “Information Effects in Collective Preferences,” American Political Science

Review, 1998, 92 (3), 545–558.

Aragones, Enriqueta and Thomas Palfrey, “Mixed Equilibrium in a Downsian Model with a

Favored Candidate,” Journal of Economic Theory, 2002, 103, 131–161.

Bernhardt, Dan, John Duggan, and Francesco Squintani, “Electoral competition with

privately-informed candidates,” Games and Economic Behavior, January 2007, 58 (1), 1–29.

, , and , “Private polling in elections and voter welfare,” Journal of Economic Theory,

September 2009, 144 (5), 2021–2056.

Besley, Tim and Stephen Coate, “An Economic Model of Representative Democracy,” Quarterly

Journal of Economics, 1997, 112, 85–114.

Bruhn, Kathleen and Kenneth F. Greene, “Elite Polarization Meets Mass Moderation in

Mexico’s 2006 Elections,” Political Science and Politics, January 2007, 39 (5), 33–38.

Callander, Steven and Simon Wilkie, “Lies, Damned Lies, and Political Campaigns,” Games

and Economic Behavior, 2007, 60 (2), 262–286.

Calvert, Randall L., “Robustness of the multidimensional voting model: Candidate motivations,

uncertainty, and convergence,” American Journal of Political Science, 1985, 29, 69–95.

Campante, Filipe R., “Redistribution in a model of voting and campaign contributions,” Journal

of Public Economics, 2011, 95 (7), 646–656.

Campbell, Angus, Philip E. Converse, Warren E. Miller, and Donald E. Stokes, The

American Voter, Chicago: University of Chicago Press, 1960.

Canes-Wrone, Brandice, Michael Herron, and Kenneth W. Shotts, “Leadership and Pan-

dering: A Theory of Executive Policymaking,” American Journal of Political Science, July 2001,

45 (3), 532–550.

Chan, Jimmy, “Electoral Competition with Private Information,” 2001. unpublished.

Che, Yeon-Koo, Wouter Dessein, and Navin Kartik, “Pandering to Persuade,” American

Economic Review, February 2013, 103 (1), 47–79.

Cukierman, Alex and Mariano Tommasi, “When Does it Take a Nixon to Go to China,”

American Economic Review, March 1998, 88 (1), 180–197.

Degroot, Morris H., Optimal Statistical Decisions, New York: McGraw-Hill, 1970.

37

Downs, Anthony, An Economic Theory of Democracy, New York: Harper and Row, 1957.

Farrell, Joseph, “Meaning and Credibility in Cheap-Talk Games,” Games and Economic Behavior,

1993, 5 (1), 514–531.

Fishkin, James S., The Voice of the People: Public Opinion and Democracy, New Haven: Yale

University Press, 1997.

Forges, Francoise, “Correlated Equilibrium in Two-Person Zero-Sum Games,” Econometrica,

March 1990, 58 (2), 515.

Gilens, Martin, “Political Ignorance and Collective Policy Preferences,” American Political Science

Review, 2001, 95 (2), 379–396.

Glaeser, Edward L. and Cass R. Sunstein, “Extremism and Social Learning,” Journal of Legal

Analysis, Winter 2009, 1, 263–324.

, Giacomo A. M. Ponzetto, and Jesse M. Shapiro, “Strategic Extremism: Why Republicans

and Democrats Divide on Religious Values,” Quarterly Journal of Economics, November 2005, 120

(4), 1283–1330.

Gratton, Gabriele, “Pandering and Electoral Competition,” Games and Economic Behavior, 2014,

84, 163–179.

Groseclose, Timothy, “A Model of Candidate Location when One Candidate Has a Valence Ad-

vantage,” American Journal of Political Science, 2001, 45, 862–886.

Harrington, Joseph E., “The Revelation of Information through the Electoral Process: An Ex-

ploratory Analysis,” Economics & Politics, 1992, 4 (3), 255–276.

, “Economic Policy, Economic Performance, and Elections,” American Economic Review, March

1993, 83 (1), 27–42.

, “The Impact of Reelection Pressures on the Fulfillment of Campaign Promises,” Games and

Economic Behavior, January 1993, 5 (1), 71–97.

Heidhues, Paul and Johan Lagerlof, “Hiding Information in Electoral Competition,” Games

and Economic Behavior, January 2003, 42 (1), 48–74.

Honryo, Takakazu, “Signaling Competence in Elections,” May 2014. working paper, University of

Mannheim.

Hotelling, Harold, “Stability in competition,” Economic Journal, March 1929, 39 (153), 41–57.

Iyengar, Shanto and Donald Kinder, News That Matters: Television and American Opinion,

Chicago: University of Chicago Press, 1987.

Jewel, W.S., “Credible Means are Exact Bayesian for Exponential Families,” ASTIN Bulletin,

1974, 8, 77–90.

Kartik, Navin and R. Preston McAfee, “Signaling Character in Electoral Competition,” Amer-

ican Economic Review, June 2007, 97 (3), 852–870.

38

and Richard van Weelden, “Informative Cheap Talk in Elections,” March 2015. unpublished.

Kass, Rob, Dennis Dannenburg, and Marc Goovaerts, “Exact Credibility for Weighted Ob-

servations,” ASTIN Bulletin, 1997, 27, 287–295.

Klumpp, Tilman, “Populism, Partisanship, and the Funding of Political Campaigns,” August

2014. unpublished.

Laslier, Jean-Francois and Karine Van de Straeten, “Electoral competition under imperfect

information,” Economic Theory, August 2004, 24 (2), 419–446.

Levy, Gilat, “Anti-herding and strategic consultation,” European Economic Review, June 2004, 48

(3), 503–525.

Loertscher, Simon, “Location Choice and Information Transmission,” 2012. mimeo, University of

Melbourne.

Majumdar, Sumon and Sharun W. Mukand, “Policy Gambles,” American Economic Review,

September 2004, 94 (4), 1207–1222.

Martinelli, Cesar, “Elections with Privately Informed Parties and Voters,” Public Choice, July

2001, 108 (1–2), 147–67.

and Akihiko Matsui, “Policy Reversals and Electoral Competition with Privately Informed

Parties,” Journal of Public Economic Theory, 2002, 4 (1), 39–61.

Maskin, Eric and Jean Tirole, “The Politician and the Judge: Accountability in Government,”

American Economic Review, September 2004, 94 (4), 1034–1054.

Matthews, Steven A., Masahiro Okuno-Fujiwara, and Andrew Postlewaite, “Refining

Cheap-Talk Equilibria,” Journal of Economic Theory, December 1991, 55 (2), 247–273.

Morelli, Massimo and Richard van Weelden, “Ideology and Information in Policy Making,”

Journal of Theoretical Politics, 2013, 25 (3), 412–439.

Myerson, Roger, “Credible negotiation statements and coherent plans,” Journal of Economic

Theory, 1989, 48 (1), 264–303.

Osborne, Martin J. and Al Slivinski, “A Model of Political Competition with Citizen-

Candidates,” Quarterly Journal of Economics, 1996, 111, 65–96.

Ottaviani, Marco and Peter Norman Sorensen, “Reputational Cheap Talk,” RAND Journal

of Economics, Spring 2006, 37 (1), 155–175.

Palfrey, Thomas, “Spatial Equilibrium with Entry,” Review of Economic Studies, 1984, 51, 139–

156.

Panova, Elena, “Informative Campaign Promises,” 2009. unpublished.

Petry, Francois and Benoıt Collette, “Measuring how political parties keep their promises: a

positive perspective from political science,” in Louis M. Imbeau, ed., Do They Walk Like They

Talk? Speech and Action in Policy Processes, Springer, 2009, chapter 5, pp. 65–80.

39

Prat, Andrea, “The Wrong Kind of Transparency,” American Economic Review, 2005, 95 (3),

862–877.

Prendergast, Canice and Lars Stole, “Impetuous Youngsters and Jaded Old-Timers: Acquiring

a Reputation for Learning,” Journal of Political Economy, December 1996, 104 (6), 1105–34.

Roux, Nicholas and Joel Sobel, “Group Polarization in a Model of Information Aggregation,”

November 2013. forthcoming in AEJ: Micro.

Schnakenberg, Keith, “Directional Cheap Talk in Electoral Campaigns,” March 2014. unpub-

lished.

Schultz, Christian, “Polarization and Inefficient Policies,” Review of Economic Studies, April 1996,

63 (2), 331–44.

Schuman, Howard and Stanley Presser, Questions and Answers in Survey Questions, San

Diego: Sage Publications, 1981.

Sørensen, Rune J., “Party polarization and government efficiency,” 2012. working paper, BI

Norwegian Business School.

Stone, Walter J. and Elizabeth N. Simas, “Candidate Valence and Ideological Positions in U.S.

House Elections,” American Journal of Political Science, April 2010, 54 (2), 371–388.

Viossat, Yannick, “The Geometry of Nash Equilibria and Correlated Equilibria and a Gener-

alization of Zero-Sum Games,” SSE/EFI Working Paper Series in Economics and Finance 641,

Stockholm School of Economics August 2006.

Wittman, Donald, “Candidates with Policy Preferences,” Journal of Economic Theory, 1977, 14,

180–189.

, “Candidate Motivation: A Synthesis of Alternatives,” American Political Science Review, 1983,

77, 142–157.

, “Why democracies produce efficient results,” Journal of Political Economy, 1989, 97, 1395–1424.

Zaller, John R., The Nature and Origins of Mass Opinion, New York: Cambridge University Press,

1992.

40

Supplementary Appendix (For Online Publication Only)

F. Proofs for Mixed Motives

Here we provide proofs for Proposition 5 and Theorem 4 from Appendix C. The following Lemma

will be used in the proof of Proposition 5.

Lemma 6. Fix any equilibrium in which candidate i is always elected, plays a pure strategy yi(·)that is a continuous function and fully revealing. Then, for any on-path platform of candidate −i,y ∈ range[yi(·)], it holds that E [θ|yi = y−i = y] = y.

Proof. Suppose, to contradiction, that E [θ|yi = y−i = y] > y for some on-path platform of −i, y ∈range [yi (·)]; the case of reverse inequality is analogous. Since yi(θi) is continuous and fully revealing,

E [θ|yi = y − ε, y−i = y] is continuous in ε. Thus, for small enough ε > 0, E [θ|yi = y − ε, y−i = y] >

y. It follows that for small enough ε > 0, the voter must elect candidate −i upon seeing yi = y − εand y−i = y. This contradicts the hypothesis that i is always elected.

Proof of Proposition 5. Without loss of generality, let i = A. Given the strategies (21) and (22),

it follows that

E[θ|yA, yB] =β(yA − bA)α+β

β + β(yB + α+β

β bA

)α+ 2β

=αyA + β(yA + yB)

α+ 2β.

Straightforward algebra then verifies that for any yA and yB,

(yA − E[θ|yA, yB])2 < (yB − E[θ|yA, yB])2 ⇐⇒ β < α+ β.

Hence it is optimal for the voter to always elect candidate A; clearly the candidates are playing

optimally given this strategy for the voter.

To prove the uniqueness claim, fix any fully revealing equilibrium in which i always wins and

plays (21). For any on-path platform of candidate −i, say y, let θ−i(y) denote the unique type that

uses platform y. Since Lemma 6 holds continues to apply when candidates have mixed motivations,

it follows that for any on-path platform y of candidate −i,

E [θ|yi = y−i = y] =β

α+ 2β

(α+ β

β(y − bi) + θ−i(y)

)= y.

Rearranging and simplifying yields θ−i(y) = y + α+ββ bi. Since this is true for any on-path platform

of −i, candidate −i must be using the pure strategy y−i (θ−i) = θ−i − α+ββ bi.

We next derive a sequence of lemmas that are needed to prove Theorem 4. We begin with

some notation. Recall that Σε(ρ, b) is the set of equilibria where each candidate wins with ex-ante

probability at least ε. Define

Σ∗(ρ, b) := σ ∈ Σ0(ρ, b) : UV (σ) = U0V (ρ, b)

41

as the set of voter-welfare-maximizing equilibria given (ρ, b).34 Say that a sequence of strategy

profiles σn → σ if: (1) for each i and θi, yni (θi) → yi(θi); and (2) for each pair (yA, yB) ∈ R2,

vn(yA, yB) → v(yA, yB).35 In other words, convergence of strategies is point-wise. Despite using

point-wise convergence, observe that because the ex-ante probability of θi : θi /∈ [−k, k] can be made

arbitrarily small by choosing k > 0 arbitrarily large, it follows that if σn → σ then UV (σn)→ UV (σ).

Lemma 7. For any (ρ, b), there is an equilibrium where both candidates play yi(·) = 0.

Proof. Immediate.

Given a strategy profile σ, let W σ(θA, θB) denote the set of candidates who win positive proba-

bility when the signal realizations are θA, θB; note that given σ, this is independent of (ρ, b).

Lemma 8. For any (θA, θB), there exists k > 0 such that for any (ρ, b), if σ ∈ Σ∗(ρ, b) and

i ∈W σ(θA, θB), then |yi(θi)| < k.36

Proof. Lemma 7 implies that for any (ρ, b), U0V (ρ, b) > −V ar(θ) = −1/α. Now fix any (θA, θB)

and note that E[θ|θA, θB] does not depend on (ρ, b). Hence, for any x > 0, there exists k > 0 such

that for any σ, if i ∈W σ(θA, θB) and |yi(θi)| > k, then the voter’s utility from σ conditional on the

realization of (θA, θB) is less than −x (using the fact that the range of v(·) is 0, 1/2, 1). Since the

voter’s utility conditional on any signal profile is bounded above by zero, it follows that there is some

x > 0 such that the voter’s utility from σ conditional on (σA, σB) being realized cannot be less than

−x if σ ∈ Σ∗(ρ), no matter what (ρ, b) is. The desired conclusion now follows.

Lemma 9. Fix any sequence of voter-welfare-maximizing equilibria as (ρ, b)→ (0,0), σρ,b ∈ Σ∗(ρ, b).

Then either:

(1) for some i, Pr(i wins in σρ,b)→ 0 as (ρ, b)→ 0; or

(2) for any i and θi, there exists k > 0 such that |yρ,bi (θi)| < k.

Proof. Suppose the lemma is false. Then, without loss, there is a type θA, a number δ > 0, and a

(sub)sequence of (ρ, b) → (0,0) with equilibria σρ,b ∈ Σ∗(ρ, b) such that: (i) for all (ρ, b) and i ∈A,B, Pr(i wins in σρ,b) > δ; and (ii) either yρ,bA (θA)→ +∞ or yρ,bA (θA)→ −∞. Lemma 8 implies

for any k > 0, there exists (ρ, b) > 0 such for any (ρ, b) < (ρ, b) if |θB| < k then A /∈W σρ,b(θA, θB).

(Intuitively, as (ρ, b) → 0, since yρ,bA (θA) explodes, it must be that type θA wins only against at

most a set of θB’s that have vanishing prior probability.) Since the distribution of θB|θA does not

change with (ρ, b), it follows that

for any ε > 0, if (ρ, b) is small enough then UA(θA;σρ,ρ, b) < ε, (26)

34 In what follows, we will proceed as if Σ∗(ρ, b) is non-empty for all (ρ, b). If this is not the case, one can proceedalmost identically, just by defining for any ε > 0, Σ∗ε(ρ, b) := σ ∈ Σ0(ρ, b) : UV (σ) ≥ U∗V (ρ, b)−ε, and then applyingthe subsequent arguments for a sequence of ε→ 0.

35 Here, yni and vn are the components of σn and similarly for the limit; a similar convention is used subsequently.Note that this supposes that candidates are playing pure strategies in equilibrium; this is for notational simplicity only,as it can be verified that the arguments go through for equilibria in which candidates may mix, with the notion ofconvergence being that of the weak topology.

36 Recall that we suppress “almost all” qualifiers.

42

where UA(θA;σ,m) is the expected utility for candidate A when his type is θA in an equilibrium σ

given candidate motivations (ρ, b). However, notice that by point (i) above, it must be that there is

a bounded set, say ΘA ⊂ R, such that for any (ρ, b), Pr(i wins in σρ,b|θA ∈ ΘA) is bounded below

by some positive number.37 But then, type θA can mimic the play of types in ΘA (e.g. mix uniformly

over their strategies) to get a strictly positive probability of winning for all (ρ, b), which given (26)

would be a profitable deviation for small enough (ρ, b).

Proof of Theorem 4. Recall that σB(ρ, b) is the equilibrium identified in Proposition 5 where,

without loss, A is the candidate who wins with probability one. We prove each part of Theorem 4

in turn.

Part 1. Let σρ,b ∈ Σ∗(ρ, b) be an arbitrary sequence of voter-welfare-maximizing equilibria as

(ρ, b)→ (0,0).38 Applying Lemma 9 to this sequence, there are two exhaustive cases:

(a) If Case 1 of Lemma 9 holds, then it is straightforward to verify that UV (σρ,b) → U0V (0,0).

Intuitively, if i is winning with ex-ante probability approximately zero, then the voter’s welfare cannot

be much higher than if −i wins with ex-ante probability one using the unbiased strategy.

(b) If Case 2 of Lemma 9 holds, pick any subsequence of σρ,b that converges (at least one

exists) and denote the limit by σ0,0. Since payoffs are continuous, it can be verified using standard

arguments that σ0,0 is an equilibrium of the limit pure-office-motivation game (intuitively, if any

type of a candidate or the voter has a profitable deviation, there would also have been a profitable

deviation from σρ,b for small enough (ρ, b) > (0,0); just as one argues in the proof of the Theorem

of the Maximum). This implies that

lim(ρ,b)→(0,0)

UV (σρ,b) = UV (σ0,0) ≤ U0V (0,0).

Since UV (σρ,b) ≥ UV (σB(ρ, b)) for all (ρ, b), it follows from Proposition 5 that in fact

lim(ρ,b)→(0,0)

UV (σρ,b) = UV (σ0,0) = U0V (0,0).

Part 2. Suppose the statement is false. Then, in light of the first part proved above, there is a

sequence of equilibria σρ,b as (ρ, b)→ (0,0) such that UV (σρ,b)→ U0V (0,0) and

for all ε > 0, a subsequence (ρ, b)ε → (0,0) where σ(ρ,b)ε ∈ Σε (ρ, b).39 (27)

Applying Lemma 9, it follows that for any ε > 0, i, and θi, there exists k > 0 such that |y(ρ,b)εi (θi)| < k.

But then, as in the first part above, σρ,b must converge (in subsequence) to some σ0,0. Since

UV (σρ,b) → U0V (0,0), it follows that σ0,0 must be non-competitive. But this implies that that for

any ε > 0, there exists δ > 0 such that if (ρ, b) is in a δ-neighborhood of (0,0), then σρ,b /∈ Σε (ρ, b),

which contradicts (27).

37 The reason ΘA must be a bounded set is because types in the tails have vanishing prior probability.

38 The same caveat as in fn. 34 applies.

39 Recall that Σε(ρ, b) is the set of equilibria where each candidate wins with ex-ante probability at least ε.

43